Skip to content

P
PaddleDetection

s920243400 / PaddleDetection
与 Fork 源项目一致

Fork自 PaddlePaddle / PaddleDetection

通知 2
Star 0
Fork 0
P
PaddleDetection
“590ecba30576f166579ac49e4fa042af9e191fd1”上不存在“paddle/legacy/function/CMakeLists.txt”
提交 fd9b650d 编写于 作者: Q qiaolongfei
浏览文件
操作

Merge branch 'develop' of https://github.com/PaddlePaddle/Paddle into update-api-reference-1

上级 d76f8a8f 7c426be9
显示空白变更内容
内联 并排
Showing with 927 addition and 519 deletion
doc/fluid/api/gen_doc.sh
浏览文件 @ fd9b650d
#!/bin/bash #!/bin/bash
python gen_doc.py layers --submodules control_flow device io nn ops tensor > layers.rst python gen_doc.py layers --submodules control_flow device io nn ops tensor detection > layers.rst
for module in data_feeder clip metrics executor initializer io nets optimizer param_attr profiler regularizer for module in data_feeder clip metrics executor initializer io nets optimizer param_attr profiler regularizer
do do
......
doc/fluid/api/io.rst
浏览文件 @ fd9b650d
...@@ -59,21 +59,3 @@ get_inference_program ...@@ -59,21 +59,3 @@ get_inference_program
.. autofunction:: paddle.fluid.io.get_inference_program .. autofunction:: paddle.fluid.io.get_inference_program
:noindex: :noindex:
save_checkpoint
---------------
.. autofunction:: paddle.fluid.io.save_checkpoint
:noindex:
load_checkpoint
---------------
.. autofunction:: paddle.fluid.io.load_checkpoint
:noindex:
clean_checkpoint
----------------
.. autofunction:: paddle.fluid.io.clean_checkpoint
:noindex:
doc/fluid/api/layers.rst
浏览文件 @ fd9b650d
...@@ -181,12 +181,6 @@ Print ...@@ -181,12 +181,6 @@ Print
.. autofunction:: paddle.fluid.layers.Print .. autofunction:: paddle.fluid.layers.Print
:noindex: :noindex:
is_empty
--------
.. autofunction:: paddle.fluid.layers.is_empty
:noindex:
device device
====== ======
...@@ -261,19 +255,6 @@ double_buffer ...@@ -261,19 +255,6 @@ double_buffer
.. autofunction:: paddle.fluid.layers.double_buffer .. autofunction:: paddle.fluid.layers.double_buffer
:noindex: :noindex:
random_data_generator
---------------------
.. autofunction:: paddle.fluid.layers.random_data_generator
:noindex:
Preprocessor
------------
.. autoclass:: paddle.fluid.layers.Preprocessor
:members:
:noindex:
nn nn
== ==
...@@ -613,30 +594,6 @@ roi_pool ...@@ -613,30 +594,6 @@ roi_pool
.. autofunction:: paddle.fluid.layers.roi_pool .. autofunction:: paddle.fluid.layers.roi_pool
:noindex: :noindex:
dice_loss
---------
.. autofunction:: paddle.fluid.layers.dice_loss
:noindex:
resize_bilinear
---------------
.. autofunction:: paddle.fluid.layers.resize_bilinear
:noindex:
gather
------
.. autofunction:: paddle.fluid.layers.gather
:noindex:
random_crop
-----------
.. autofunction:: paddle.fluid.layers.random_crop
:noindex:
ops ops
=== ===
...@@ -784,12 +741,6 @@ sum ...@@ -784,12 +741,6 @@ sum
.. autofunction:: paddle.fluid.layers.sum .. autofunction:: paddle.fluid.layers.sum
:noindex: :noindex:
shape
-----
.. autofunction:: paddle.fluid.layers.shape
:noindex:
sigmoid sigmoid
------- -------
...@@ -1039,3 +990,54 @@ zeros ...@@ -1039,3 +990,54 @@ zeros
.. autofunction:: paddle.fluid.layers.zeros .. autofunction:: paddle.fluid.layers.zeros
:noindex: :noindex:
detection
=========
multi_box_head
--------------
.. autofunction:: paddle.fluid.layers.multi_box_head
:noindex:
bipartite_match
---------------
.. autofunction:: paddle.fluid.layers.bipartite_match
:noindex:
target_assign
-------------
.. autofunction:: paddle.fluid.layers.target_assign
:noindex:
detection_output
----------------
.. autofunction:: paddle.fluid.layers.detection_output
:noindex:
ssd_loss
--------
.. autofunction:: paddle.fluid.layers.ssd_loss
:noindex:
detection_map
-------------
.. autofunction:: paddle.fluid.layers.detection_map
:noindex:
iou_similarity
--------------
.. autofunction:: paddle.fluid.layers.iou_similarity
:noindex:
box_coder
---------
.. autofunction:: paddle.fluid.layers.box_coder
:noindex:
doc/fluid/api/optimizer.rst
浏览文件 @ fd9b650d
...@@ -89,13 +89,6 @@ DecayedAdagradOptimizer ...@@ -89,13 +89,6 @@ DecayedAdagradOptimizer
:members: :members:
:noindex: :noindex:
RMSPropOptimizer
----------------
.. autoclass:: paddle.fluid.optimizer.RMSPropOptimizer
:members:
:noindex:
Adadelta Adadelta
-------- --------
......
doc/fluid/api/profiler.rst
浏览文件 @ fd9b650d
...@@ -23,15 +23,3 @@ profiler ...@@ -23,15 +23,3 @@ profiler
.. autofunction:: paddle.fluid.profiler.profiler .. autofunction:: paddle.fluid.profiler.profiler
:noindex: :noindex:
start_profiler
--------------
.. autofunction:: paddle.fluid.profiler.start_profiler
:noindex:
stop_profiler
-------------
.. autofunction:: paddle.fluid.profiler.stop_profiler
:noindex:
doc/v2/api/config/evaluators.rst
浏览文件 @ fd9b650d
...@@ -101,7 +101,7 @@ value_printer ...@@ -101,7 +101,7 @@ value_printer
:noindex: :noindex:
Detection Detection
===== ==========
detection_map detection_map
------------- -------------
......
doc/v2/api/config/layer.rst
浏览文件 @ fd9b650d
...@@ -11,7 +11,7 @@ Data layer ...@@ -11,7 +11,7 @@ Data layer
data data
---- ----
.. autoclass:: paddle.v2.layer.data .. autofunction:: paddle.v2.layer.data
:noindex: :noindex:
Fully Connected Layers Fully Connected Layers
...@@ -21,12 +21,12 @@ Fully Connected Layers ...@@ -21,12 +21,12 @@ Fully Connected Layers
fc fc
-- --
.. autoclass:: paddle.v2.layer.fc .. autofunction:: paddle.v2.layer.fc
:noindex: :noindex:
selective_fc selective_fc
------------ ------------
.. autoclass:: paddle.v2.layer.selective_fc .. autofunction:: paddle.v2.layer.selective_fc
:noindex: :noindex:
Conv Layers Conv Layers
...@@ -34,34 +34,34 @@ Conv Layers ...@@ -34,34 +34,34 @@ Conv Layers
conv_operator conv_operator
------------- -------------
.. autoclass:: paddle.v2.layer.conv_operator .. autofunction:: paddle.v2.layer.conv_operator
:noindex: :noindex:
conv_projection conv_projection
--------------- ---------------
.. autoclass:: paddle.v2.layer.conv_projection .. autofunction:: paddle.v2.layer.conv_projection
:noindex: :noindex:
conv_shift conv_shift
---------- ----------
.. autoclass:: paddle.v2.layer.conv_shift .. autofunction:: paddle.v2.layer.conv_shift
:noindex: :noindex:
img_conv img_conv
-------- --------
.. autoclass:: paddle.v2.layer.img_conv .. autofunction:: paddle.v2.layer.img_conv
:noindex: :noindex:
.. _api_v2.layer_context_projection: .. _api_v2.layer_context_projection:
context_projection context_projection
------------------ ------------------
.. autoclass:: paddle.v2.layer.context_projection .. autofunction:: paddle.v2.layer.context_projection
:noindex: :noindex:
row_conv row_conv
-------- --------
.. autoclass:: paddle.v2.layer.row_conv .. autofunction:: paddle.v2.layer.row_conv
:noindex: :noindex:
Image Pooling Layer Image Pooling Layer
...@@ -69,27 +69,27 @@ Image Pooling Layer ...@@ -69,27 +69,27 @@ Image Pooling Layer
img_pool img_pool
-------- --------
.. autoclass:: paddle.v2.layer.img_pool .. autofunction:: paddle.v2.layer.img_pool
:noindex: :noindex:
spp spp
--- ---
.. autoclass:: paddle.v2.layer.spp .. autofunction:: paddle.v2.layer.spp
:noindex: :noindex:
maxout maxout
------ ------
.. autoclass:: paddle.v2.layer.maxout .. autofunction:: paddle.v2.layer.maxout
:noindex: :noindex:
roi_pool roi_pool
-------- --------
.. autoclass:: paddle.v2.layer.roi_pool .. autofunction:: paddle.v2.layer.roi_pool
:noindex: :noindex:
pad pad
---- ----
.. autoclass:: paddle.v2.layer.pad .. autofunction:: paddle.v2.layer.pad
:noindex: :noindex:
Norm Layer Norm Layer
...@@ -97,27 +97,27 @@ Norm Layer ...@@ -97,27 +97,27 @@ Norm Layer
img_cmrnorm img_cmrnorm
----------- -----------
.. autoclass:: paddle.v2.layer.img_cmrnorm .. autofunction:: paddle.v2.layer.img_cmrnorm
:noindex: :noindex:
batch_norm batch_norm
---------- ----------
.. autoclass:: paddle.v2.layer.batch_norm .. autofunction:: paddle.v2.layer.batch_norm
:noindex: :noindex:
sum_to_one_norm sum_to_one_norm
--------------- ---------------
.. autoclass:: paddle.v2.layer.sum_to_one_norm .. autofunction:: paddle.v2.layer.sum_to_one_norm
:noindex: :noindex:
cross_channel_norm cross_channel_norm
------------------ ------------------
.. autoclass:: paddle.v2.layer.cross_channel_norm .. autofunction:: paddle.v2.layer.cross_channel_norm
:noindex: :noindex:
row_l2_norm row_l2_norm
----------- -----------
.. autoclass:: paddle.v2.layer.row_l2_norm .. autofunction:: paddle.v2.layer.row_l2_norm
:noindex: :noindex:
Recurrent Layers Recurrent Layers
...@@ -125,22 +125,22 @@ Recurrent Layers ...@@ -125,22 +125,22 @@ Recurrent Layers
recurrent recurrent
--------- ---------
.. autoclass:: paddle.v2.layer.recurrent .. autofunction:: paddle.v2.layer.recurrent
:noindex: :noindex:
lstmemory lstmemory
--------- ---------
.. autoclass:: paddle.v2.layer.lstmemory .. autofunction:: paddle.v2.layer.lstmemory
:noindex: :noindex:
grumemory grumemory
--------- ---------
.. autoclass:: paddle.v2.layer.grumemory .. autofunction:: paddle.v2.layer.grumemory
:noindex: :noindex:
gated_unit gated_unit
----------- -----------
.. autoclass:: paddle.v2.layer.gated_unit .. autofunction:: paddle.v2.layer.gated_unit
:noindex: :noindex:
Recurrent Layer Group Recurrent Layer Group
...@@ -148,32 +148,32 @@ Recurrent Layer Group ...@@ -148,32 +148,32 @@ Recurrent Layer Group
memory memory
------ ------
.. autoclass:: paddle.v2.layer.memory .. autofunction:: paddle.v2.layer.memory
:noindex: :noindex:
recurrent_group recurrent_group
--------------- ---------------
.. autoclass:: paddle.v2.layer.recurrent_group .. autofunction:: paddle.v2.layer.recurrent_group
:noindex: :noindex:
lstm_step lstm_step
--------- ---------
.. autoclass:: paddle.v2.layer.lstm_step .. autofunction:: paddle.v2.layer.lstm_step
:noindex: :noindex:
gru_step gru_step
-------- --------
.. autoclass:: paddle.v2.layer.gru_step .. autofunction:: paddle.v2.layer.gru_step
:noindex: :noindex:
beam_search beam_search
------------ ------------
.. autoclass:: paddle.v2.layer.beam_search .. autofunction:: paddle.v2.layer.beam_search
:noindex: :noindex:
get_output get_output
---------- ----------
.. autoclass:: paddle.v2.layer.get_output .. autofunction:: paddle.v2.layer.get_output
:noindex: :noindex:
Mixed Layer Mixed Layer
...@@ -183,54 +183,54 @@ Mixed Layer ...@@ -183,54 +183,54 @@ Mixed Layer
mixed mixed
----- -----
.. autoclass:: paddle.v2.layer.mixed .. autofunction:: paddle.v2.layer.mixed
:noindex: :noindex:
.. _api_v2.layer_embedding: .. _api_v2.layer_embedding:
embedding embedding
--------- ---------
.. autoclass:: paddle.v2.layer.embedding .. autofunction:: paddle.v2.layer.embedding
:noindex: :noindex:
scaling_projection scaling_projection
------------------ ------------------
.. autoclass:: paddle.v2.layer.scaling_projection .. autofunction:: paddle.v2.layer.scaling_projection
:noindex: :noindex:
dotmul_projection dotmul_projection
----------------- -----------------
.. autoclass:: paddle.v2.layer.dotmul_projection .. autofunction:: paddle.v2.layer.dotmul_projection
:noindex: :noindex:
dotmul_operator dotmul_operator
--------------- ---------------
.. autoclass:: paddle.v2.layer.dotmul_operator .. autofunction:: paddle.v2.layer.dotmul_operator
:noindex: :noindex:
full_matrix_projection full_matrix_projection
---------------------- ----------------------
.. autoclass:: paddle.v2.layer.full_matrix_projection .. autofunction:: paddle.v2.layer.full_matrix_projection
:noindex: :noindex:
identity_projection identity_projection
------------------- -------------------
.. autoclass:: paddle.v2.layer.identity_projection .. autofunction:: paddle.v2.layer.identity_projection
:noindex: :noindex:
slice_projection slice_projection
------------------- -------------------
.. autoclass:: paddle.v2.layer.slice_projection .. autofunction:: paddle.v2.layer.slice_projection
:noindex: :noindex:
table_projection table_projection
---------------- ----------------
.. autoclass:: paddle.v2.layer.table_projection .. autofunction:: paddle.v2.layer.table_projection
:noindex: :noindex:
trans_full_matrix_projection trans_full_matrix_projection
---------------------------- ----------------------------
.. autoclass:: paddle.v2.layer.trans_full_matrix_projection .. autofunction:: paddle.v2.layer.trans_full_matrix_projection
:noindex: :noindex:
Aggregate Layers Aggregate Layers
...@@ -245,51 +245,46 @@ AggregateLevel ...@@ -245,51 +245,46 @@ AggregateLevel
pooling pooling
------- -------
.. autoclass:: paddle.v2.layer.pooling .. autofunction:: paddle.v2.layer.pooling
:noindex: :noindex:
.. _api_v2.layer_last_seq: .. _api_v2.layer_last_seq:
last_seq last_seq
-------- --------
.. autoclass:: paddle.v2.layer.last_seq .. autofunction:: paddle.v2.layer.last_seq
:noindex: :noindex:
.. _api_v2.layer_first_seq: .. _api_v2.layer_first_seq:
first_seq first_seq
--------- ---------
.. autoclass:: paddle.v2.layer.first_seq .. autofunction:: paddle.v2.layer.first_seq
:noindex: :noindex:
sub_seq sub_seq
--------- ---------
.. autoclass:: paddle.v2.layer.sub_seq .. autofunction:: paddle.v2.layer.sub_seq
:noindex: :noindex:
concat concat
------ ------
.. autoclass:: paddle.v2.layer.concat .. autofunction:: paddle.v2.layer.concat
:noindex: :noindex:
seq_concat seq_concat
---------- ----------
.. autoclass:: paddle.v2.layer.seq_concat .. autofunction:: paddle.v2.layer.seq_concat
:noindex: :noindex:
seq_slice seq_slice
--------- ---------
.. autoclass:: paddle.v2.layer.seq_slice .. autofunction:: paddle.v2.layer.seq_slice
:noindex:
kmax_sequence_score
-------------------
.. autoclass:: paddle.v2.layer.kmax_sequence_score
:noindex: :noindex:
sub_nested_seq sub_nested_seq
-------------- --------------
.. autoclass:: paddle.v2.layer.sub_nested_seq .. autofunction:: paddle.v2.layer.sub_nested_seq
:noindex: :noindex:
Reshaping Layers Reshaping Layers
...@@ -297,7 +292,7 @@ Reshaping Layers ...@@ -297,7 +292,7 @@ Reshaping Layers
block_expand block_expand
------------ ------------
.. autoclass:: paddle.v2.layer.block_expand .. autofunction:: paddle.v2.layer.block_expand
:noindex: :noindex:
.. _api_v2.layer_expand: .. _api_v2.layer_expand:
...@@ -309,22 +304,22 @@ ExpandLevel ...@@ -309,22 +304,22 @@ ExpandLevel
expand expand
------ ------
.. autoclass:: paddle.v2.layer.expand .. autofunction:: paddle.v2.layer.expand
:noindex: :noindex:
repeat repeat
------ ------
.. autoclass:: paddle.v2.layer.repeat .. autofunction:: paddle.v2.layer.repeat
:noindex: :noindex:
rotate rotate
------ ------
.. autoclass:: paddle.v2.layer.rotate .. autofunction:: paddle.v2.layer.rotate
:noindex: :noindex:
seq_reshape seq_reshape
----------- -----------
.. autoclass:: paddle.v2.layer.seq_reshape .. autofunction:: paddle.v2.layer.seq_reshape
:noindex: :noindex:
Math Layers Math Layers
...@@ -332,94 +327,94 @@ Math Layers ...@@ -332,94 +327,94 @@ Math Layers
addto addto
----- -----
.. autoclass:: paddle.v2.layer.addto .. autofunction:: paddle.v2.layer.addto
:noindex: :noindex:
linear_comb linear_comb
----------- -----------
.. autoclass:: paddle.v2.layer.linear_comb .. autofunction:: paddle.v2.layer.linear_comb
:noindex: :noindex:
interpolation interpolation
------------- -------------
.. autoclass:: paddle.v2.layer.interpolation .. autofunction:: paddle.v2.layer.interpolation
:noindex: :noindex:
bilinear_interp bilinear_interp
--------------- ---------------
.. autoclass:: paddle.v2.layer.bilinear_interp .. autofunction:: paddle.v2.layer.bilinear_interp
:noindex: :noindex:
dropout dropout
-------- --------
.. autoclass:: paddle.v2.layer.dropout .. autofunction:: paddle.v2.layer.dropout
:noindex: :noindex:
dot_prod dot_prod
--------- ---------
.. autoclass:: paddle.v2.layer.dot_prod .. autofunction:: paddle.v2.layer.dot_prod
:noindex: :noindex:
out_prod out_prod
-------- --------
.. autoclass:: paddle.v2.layer.out_prod .. autofunction:: paddle.v2.layer.out_prod
:noindex: :noindex:
power power
----- -----
.. autoclass:: paddle.v2.layer.power .. autofunction:: paddle.v2.layer.power
:noindex: :noindex:
scaling scaling
------- -------
.. autoclass:: paddle.v2.layer.scaling .. autofunction:: paddle.v2.layer.scaling
:noindex: :noindex:
clip clip
---- ----
.. autoclass:: paddle.v2.layer.clip .. autofunction:: paddle.v2.layer.clip
:noindex: :noindex:
resize resize
------ ------
.. autoclass:: paddle.v2.layer.resize .. autofunction:: paddle.v2.layer.resize
:noindex: :noindex:
slope_intercept slope_intercept
--------------- ---------------
.. autoclass:: paddle.v2.layer.slope_intercept .. autofunction:: paddle.v2.layer.slope_intercept
:noindex: :noindex:
tensor tensor
------ ------
.. autoclass:: paddle.v2.layer.tensor .. autofunction:: paddle.v2.layer.tensor
:noindex: :noindex:
.. _api_v2.layer_cos_sim: .. _api_v2.layer_cos_sim:
cos_sim cos_sim
------- -------
.. autoclass:: paddle.v2.layer.cos_sim .. autofunction:: paddle.v2.layer.cos_sim
:noindex: :noindex:
l2_distance l2_distance
----------- -----------
.. autoclass:: paddle.v2.layer.l2_distance .. autofunction:: paddle.v2.layer.l2_distance
:noindex: :noindex:
trans trans
----- -----
.. autoclass:: paddle.v2.layer.trans .. autofunction:: paddle.v2.layer.trans
:noindex: :noindex:
scale_shift scale_shift
----------- -----------
.. autoclass:: paddle.v2.layer.scale_shift .. autofunction:: paddle.v2.layer.scale_shift
:noindex: :noindex:
factorization_machine factorization_machine
--------------------- ---------------------
.. autoclass:: paddle.v2.layer.factorization_machine .. autofunction:: paddle.v2.layer.factorization_machine
:noindex: :noindex:
Sampling Layers Sampling Layers
...@@ -427,17 +422,17 @@ Sampling Layers ...@@ -427,17 +422,17 @@ Sampling Layers
maxid maxid
----- -----
.. autoclass:: paddle.v2.layer.max_id .. autofunction:: paddle.v2.layer.max_id
:noindex: :noindex:
sampling_id sampling_id
----------- -----------
.. autoclass:: paddle.v2.layer.sampling_id .. autofunction:: paddle.v2.layer.sampling_id
:noindex: :noindex:
multiplex multiplex
--------- ---------
.. autoclass:: paddle.v2.layer.multiplex .. autofunction:: paddle.v2.layer.multiplex
:noindex: :noindex:
.. _api_v2.layer_costs: .. _api_v2.layer_costs:
...@@ -447,97 +442,97 @@ Cost Layers ...@@ -447,97 +442,97 @@ Cost Layers
cross_entropy_cost cross_entropy_cost
------------------ ------------------
.. autoclass:: paddle.v2.layer.cross_entropy_cost .. autofunction:: paddle.v2.layer.cross_entropy_cost
:noindex: :noindex:
cross_entropy_with_selfnorm_cost cross_entropy_with_selfnorm_cost
-------------------------------- --------------------------------
.. autoclass:: paddle.v2.layer.cross_entropy_with_selfnorm_cost .. autofunction:: paddle.v2.layer.cross_entropy_with_selfnorm_cost
:noindex: :noindex:
multi_binary_label_cross_entropy_cost multi_binary_label_cross_entropy_cost
------------------------------------- -------------------------------------
.. autoclass:: paddle.v2.layer.multi_binary_label_cross_entropy_cost .. autofunction:: paddle.v2.layer.multi_binary_label_cross_entropy_cost
:noindex: :noindex:
classification_cost classification_cost
------------------- -------------------
.. autoclass:: paddle.v2.layer.classification_cost .. autofunction:: paddle.v2.layer.classification_cost
:noindex: :noindex:
huber_regression_cost huber_regression_cost
------------------------- -------------------------
.. autoclass:: paddle.v2.layer.huber_regression_cost .. autofunction:: paddle.v2.layer.huber_regression_cost
:noindex: :noindex:
huber_classification_cost huber_classification_cost
------------------------- -------------------------
.. autoclass:: paddle.v2.layer.huber_classification_cost .. autofunction:: paddle.v2.layer.huber_classification_cost
:noindex: :noindex:
lambda_cost lambda_cost
----------- -----------
.. autoclass:: paddle.v2.layer.lambda_cost .. autofunction:: paddle.v2.layer.lambda_cost
:noindex: :noindex:
square_error_cost square_error_cost
----------------- -----------------
.. autoclass:: paddle.v2.layer.square_error_cost .. autofunction:: paddle.v2.layer.square_error_cost
:noindex: :noindex:
rank_cost rank_cost
--------- ---------
.. autoclass:: paddle.v2.layer.rank_cost .. autofunction:: paddle.v2.layer.rank_cost
:noindex: :noindex:
sum_cost sum_cost
--------- ---------
.. autoclass:: paddle.v2.layer.sum_cost .. autofunction:: paddle.v2.layer.sum_cost
:noindex: :noindex:
crf crf
--- ---
.. autoclass:: paddle.v2.layer.crf .. autofunction:: paddle.v2.layer.crf
:noindex: :noindex:
crf_decoding crf_decoding
------------ ------------
.. autoclass:: paddle.v2.layer.crf_decoding .. autofunction:: paddle.v2.layer.crf_decoding
:noindex: :noindex:
ctc ctc
--- ---
.. autoclass:: paddle.v2.layer.ctc .. autofunction:: paddle.v2.layer.ctc
:noindex: :noindex:
warp_ctc warp_ctc
-------- --------
.. autoclass:: paddle.v2.layer.warp_ctc .. autofunction:: paddle.v2.layer.warp_ctc
:noindex: :noindex:
nce nce
--- ---
.. autoclass:: paddle.v2.layer.nce .. autofunction:: paddle.v2.layer.nce
:noindex: :noindex:
hsigmoid hsigmoid
--------- ---------
.. autoclass:: paddle.v2.layer.hsigmoid .. autofunction:: paddle.v2.layer.hsigmoid
:noindex: :noindex:
smooth_l1_cost smooth_l1_cost
-------------- --------------
.. autoclass:: paddle.v2.layer.smooth_l1_cost .. autofunction:: paddle.v2.layer.smooth_l1_cost
:noindex: :noindex:
multibox_loss multibox_loss
-------------- --------------
.. autoclass:: paddle.v2.layer.multibox_loss .. autofunction:: paddle.v2.layer.multibox_loss
:noindex: :noindex:
detection_output detection_output
---------------- ----------------
.. autoclass:: paddle.v2.layer.detection_output .. autofunction:: paddle.v2.layer.detection_output
:noindex: :noindex:
Check Layer Check Layer
...@@ -545,7 +540,7 @@ Check Layer ...@@ -545,7 +540,7 @@ Check Layer
eos eos
--- ---
.. autoclass:: paddle.v2.layer.eos .. autofunction:: paddle.v2.layer.eos
:noindex: :noindex:
Activation Activation
...@@ -553,5 +548,5 @@ Activation ...@@ -553,5 +548,5 @@ Activation
prelu prelu
-------- --------
.. autoclass:: paddle.v2.layer.prelu .. autofunction:: paddle.v2.layer.prelu
:noindex: :noindex:
doc/v2/api/index_en.rst
浏览文件 @ fd9b650d
...@@ -8,4 +8,3 @@ API ...@@ -8,4 +8,3 @@ API
model_configs.rst model_configs.rst
data.rst data.rst
run_logic.rst run_logic.rst
fluid/index.rst
doc/v2/build_and_install/pip_install_cn.rst
浏览文件 @ fd9b650d
...@@ -60,6 +60,7 @@ paddlepaddle-gpu==0.11.0 使用CUDA 7.5和cuDNN 5编译的0.11.0版 ...@@ -60,6 +60,7 @@ paddlepaddle-gpu==0.11.0 使用CUDA 7.5和cuDNN 5编译的0.11.0版
"cpu_noavx_openblas", "`paddlepaddle-latest-cp27-cp27mu-linux_x86_64.whl <https://guest:@paddleci.ngrok.io/repository/download/Manylinux1_CpuNoavxOpenblas/.lastSuccessful/paddlepaddle-latest-cp27-cp27mu-linux_x86_64.whl>`__", "`paddlepaddle-latest-cp27-cp27m-linux_x86_64.whl <https://guest:@paddleci.ngrok.io/repository/download/Manylinux1_CpuNoavxOpenblas/.lastSuccessful/paddlepaddle-latest-cp27-cp27m-linux_x86_64.whl>`_" "cpu_noavx_openblas", "`paddlepaddle-latest-cp27-cp27mu-linux_x86_64.whl <https://guest:@paddleci.ngrok.io/repository/download/Manylinux1_CpuNoavxOpenblas/.lastSuccessful/paddlepaddle-latest-cp27-cp27mu-linux_x86_64.whl>`__", "`paddlepaddle-latest-cp27-cp27m-linux_x86_64.whl <https://guest:@paddleci.ngrok.io/repository/download/Manylinux1_CpuNoavxOpenblas/.lastSuccessful/paddlepaddle-latest-cp27-cp27m-linux_x86_64.whl>`_"
"cuda8.0_cudnn5_avx_mkl", "`paddlepaddle_gpu-latest-cp27-cp27mu-linux_x86_64.whl <https://guest:@paddleci.ngrok.io/repository/download/Manylinux1_Cuda80cudnn5cp27cp27mu/.lastSuccessful/paddlepaddle_gpu-latest-cp27-cp27mu-linux_x86_64.whl>`__", "`paddlepaddle_gpu-latest-cp27-cp27m-linux_x86_64.whl <https://guest:@paddleci.ngrok.io/repository/download/Manylinux1_Cuda80cudnn5cp27cp27mu/.lastSuccessful/paddlepaddle_gpu-latest-cp27-cp27m-linux_x86_64.whl>`__" "cuda8.0_cudnn5_avx_mkl", "`paddlepaddle_gpu-latest-cp27-cp27mu-linux_x86_64.whl <https://guest:@paddleci.ngrok.io/repository/download/Manylinux1_Cuda80cudnn5cp27cp27mu/.lastSuccessful/paddlepaddle_gpu-latest-cp27-cp27mu-linux_x86_64.whl>`__", "`paddlepaddle_gpu-latest-cp27-cp27m-linux_x86_64.whl <https://guest:@paddleci.ngrok.io/repository/download/Manylinux1_Cuda80cudnn5cp27cp27mu/.lastSuccessful/paddlepaddle_gpu-latest-cp27-cp27m-linux_x86_64.whl>`__"
"cuda8.0_cudnn7_avx_mkl", "`paddlepaddle_gpu-latest-cp27-cp27mu-linux_x86_64.whl <https://guest:@paddleci.ngrok.io/repository/download/Manylinux1_Cuda8cudnn7cp27cp27mu/.lastSuccessful/paddlepaddle_gpu-latest-cp27-cp27mu-linux_x86_64.whl>`__", "`paddlepaddle_gpu-latest-cp27-cp27m-linux_x86_64.whl <https://guest:@paddleci.ngrok.io/repository/download/Manylinux1_Cuda8cudnn7cp27cp27mu/.lastSuccessful/paddlepaddle_gpu-latest-cp27-cp27m-linux_x86_64.whl>`__" "cuda8.0_cudnn7_avx_mkl", "`paddlepaddle_gpu-latest-cp27-cp27mu-linux_x86_64.whl <https://guest:@paddleci.ngrok.io/repository/download/Manylinux1_Cuda8cudnn7cp27cp27mu/.lastSuccessful/paddlepaddle_gpu-latest-cp27-cp27mu-linux_x86_64.whl>`__", "`paddlepaddle_gpu-latest-cp27-cp27m-linux_x86_64.whl <https://guest:@paddleci.ngrok.io/repository/download/Manylinux1_Cuda8cudnn7cp27cp27mu/.lastSuccessful/paddlepaddle_gpu-latest-cp27-cp27m-linux_x86_64.whl>`__"
"cuda9.0_cudnn7_avx_mkl", "`paddlepaddle_gpu-latest-cp27-cp27mu-linux_x86_64.whl <https://guest:@paddleci.ngrok.io/repository/download/Manylinux1_Cuda90cudnn7avxMkl/.lastSuccessful/paddlepaddle_gpu-latest-cp27-cp27mu-linux_x86_64.whl>`__", "`paddlepaddle_gpu-latest-cp27-cp27m-linux_x86_64.whl <https://guest:@paddleci.ngrok.io/repository/download/Manylinux1_Cuda90cudnn7avxMkl/.lastSuccessful/paddlepaddle_gpu-latest-cp27-cp27m-linux_x86_64.whl>`__"
.. _pip_dependency: .. _pip_dependency:
......
doc/v2/build_and_install/pip_install_en.rst
浏览文件 @ fd9b650d
...@@ -63,6 +63,7 @@ If the links below shows up the login form, just click "Log in as guest" to star ...@@ -63,6 +63,7 @@ If the links below shows up the login form, just click "Log in as guest" to star
"cpu_noavx_openblas", "`paddlepaddle-latest-cp27-cp27mu-linux_x86_64.whl <https://guest:@paddleci.ngrok.io/repository/download/Manylinux1_CpuNoavxOpenblas/.lastSuccessful/paddlepaddle-latest-cp27-cp27mu-linux_x86_64.whl>`__", "`paddlepaddle-latest-cp27-cp27m-linux_x86_64.whl <https://guest:@paddleci.ngrok.io/repository/download/Manylinux1_CpuNoavxOpenblas/.lastSuccessful/paddlepaddle-latest-cp27-cp27m-linux_x86_64.whl>`__" "cpu_noavx_openblas", "`paddlepaddle-latest-cp27-cp27mu-linux_x86_64.whl <https://guest:@paddleci.ngrok.io/repository/download/Manylinux1_CpuNoavxOpenblas/.lastSuccessful/paddlepaddle-latest-cp27-cp27mu-linux_x86_64.whl>`__", "`paddlepaddle-latest-cp27-cp27m-linux_x86_64.whl <https://guest:@paddleci.ngrok.io/repository/download/Manylinux1_CpuNoavxOpenblas/.lastSuccessful/paddlepaddle-latest-cp27-cp27m-linux_x86_64.whl>`__"
"cuda8.0_cudnn5_avx_mkl", "`paddlepaddle_gpu-latest-cp27-cp27mu-linux_x86_64.whl <https://guest:@paddleci.ngrok.io/repository/download/Manylinux1_Cuda80cudnn5cp27cp27mu/.lastSuccessful/paddlepaddle_gpu-latest-cp27-cp27mu-linux_x86_64.whl>`__", "`paddlepaddle_gpu-latest-cp27-cp27m-linux_x86_64.whl <https://guest:@paddleci.ngrok.io/repository/download/Manylinux1_Cuda80cudnn5cp27cp27mu/.lastSuccessful/paddlepaddle_gpu-latest-cp27-cp27m-linux_x86_64.whl>`__" "cuda8.0_cudnn5_avx_mkl", "`paddlepaddle_gpu-latest-cp27-cp27mu-linux_x86_64.whl <https://guest:@paddleci.ngrok.io/repository/download/Manylinux1_Cuda80cudnn5cp27cp27mu/.lastSuccessful/paddlepaddle_gpu-latest-cp27-cp27mu-linux_x86_64.whl>`__", "`paddlepaddle_gpu-latest-cp27-cp27m-linux_x86_64.whl <https://guest:@paddleci.ngrok.io/repository/download/Manylinux1_Cuda80cudnn5cp27cp27mu/.lastSuccessful/paddlepaddle_gpu-latest-cp27-cp27m-linux_x86_64.whl>`__"
"cuda8.0_cudnn7_avx_mkl", "`paddlepaddle_gpu-latest-cp27-cp27mu-linux_x86_64.whl <https://guest:@paddleci.ngrok.io/repository/download/Manylinux1_Cuda8cudnn7cp27cp27mu/.lastSuccessful/paddlepaddle_gpu-latest-cp27-cp27mu-linux_x86_64.whl>`__", "`paddlepaddle_gpu-latest-cp27-cp27m-linux_x86_64.whl <https://guest:@paddleci.ngrok.io/repository/download/Manylinux1_Cuda8cudnn7cp27cp27mu/.lastSuccessful/paddlepaddle_gpu-latest-cp27-cp27m-linux_x86_64.whl>`__" "cuda8.0_cudnn7_avx_mkl", "`paddlepaddle_gpu-latest-cp27-cp27mu-linux_x86_64.whl <https://guest:@paddleci.ngrok.io/repository/download/Manylinux1_Cuda8cudnn7cp27cp27mu/.lastSuccessful/paddlepaddle_gpu-latest-cp27-cp27mu-linux_x86_64.whl>`__", "`paddlepaddle_gpu-latest-cp27-cp27m-linux_x86_64.whl <https://guest:@paddleci.ngrok.io/repository/download/Manylinux1_Cuda8cudnn7cp27cp27mu/.lastSuccessful/paddlepaddle_gpu-latest-cp27-cp27m-linux_x86_64.whl>`__"
"cuda9.0_cudnn7_avx_mkl", "`paddlepaddle_gpu-latest-cp27-cp27mu-linux_x86_64.whl <https://guest:@paddleci.ngrok.io/repository/download/Manylinux1_Cuda90cudnn7avxMkl/.lastSuccessful/paddlepaddle_gpu-latest-cp27-cp27mu-linux_x86_64.whl>`__", "`paddlepaddle_gpu-latest-cp27-cp27m-linux_x86_64.whl <https://guest:@paddleci.ngrok.io/repository/download/Manylinux1_Cuda90cudnn7avxMkl/.lastSuccessful/paddlepaddle_gpu-latest-cp27-cp27m-linux_x86_64.whl>`__"
.. _pip_dependency: .. _pip_dependency:
......
paddle/fluid/framework/CMakeLists.txt
浏览文件 @ fd9b650d
...@@ -84,7 +84,7 @@ cc_library(lod_rank_table SRCS lod_rank_table.cc DEPS lod_tensor) ...@@ -84,7 +84,7 @@ cc_library(lod_rank_table SRCS lod_rank_table.cc DEPS lod_tensor)
cc_library(feed_fetch_method SRCS feed_fetch_method.cc DEPS lod_tensor scope glog) cc_library(feed_fetch_method SRCS feed_fetch_method.cc DEPS lod_tensor scope glog)
if(WITH_DISTRIBUTE) if(WITH_DISTRIBUTE)
cc_library(executor SRCS executor.cc DEPS op_registry device_context scope framework_proto glog lod_rank_table feed_fetch_method sendrecvop_grpc grpc++_unsecure grpc_unsecure gpr) cc_library(executor SRCS executor.cc DEPS op_registry device_context scope framework_proto glog lod_rank_table feed_fetch_method sendrecvop_grpc cares grpc++_unsecure grpc_unsecure gpr)
set(DISTRIBUTE_COMPILE_FLAGS "-Wno-non-virtual-dtor -Wno-error=non-virtual-dtor -Wno-error=delete-non-virtual-dtor") set(DISTRIBUTE_COMPILE_FLAGS "-Wno-non-virtual-dtor -Wno-error=non-virtual-dtor -Wno-error=delete-non-virtual-dtor")
set_source_files_properties(executor.cc PROPERTIES COMPILE_FLAGS ${DISTRIBUTE_COMPILE_FLAGS}) set_source_files_properties(executor.cc PROPERTIES COMPILE_FLAGS ${DISTRIBUTE_COMPILE_FLAGS})
else() else()
......
paddle/fluid/framework/executor.cc
浏览文件 @ fd9b650d
...@@ -330,8 +330,12 @@ void Executor::RunPreparedContext(ExecutorPrepareContext* ctx, Scope* scope, ...@@ -330,8 +330,12 @@ void Executor::RunPreparedContext(ExecutorPrepareContext* ctx, Scope* scope,
} }
for (auto& op : ctx->ops_) { for (auto& op : ctx->ops_) {
VLOG(3) << place_ << " " << op->DebugStringEx(local_scope); VLOG(4) << place_ << " " << op->DebugStringEx(local_scope);
op->Run(*local_scope, place_); op->Run(*local_scope, place_);
// NOTE! Please do not delete this line, it's usefull because the debug
// string before and after op.run are different, after run the output
// will have right shape which is usefull for debug.
VLOG(3) << place_ << " " << op->DebugStringEx(local_scope);
if (FLAGS_benchmark) { if (FLAGS_benchmark) {
VLOG(2) << "Memory used after operator " + op->Type() + " running: " VLOG(2) << "Memory used after operator " + op->Type() + " running: "
......
paddle/fluid/framework/operator.cc
浏览文件 @ fd9b650d
...@@ -69,6 +69,19 @@ static DDim GetDims(const Scope& scope, const std::string& name, ...@@ -69,6 +69,19 @@ static DDim GetDims(const Scope& scope, const std::string& name,
} }
} }
static int GetRowSize(const Scope& scope, const std::string& name) {
Variable* var = scope.FindVar(name);
if (var == nullptr) {
return -1;
}
if (var->IsType<SelectedRows>()) {
return var->Get<SelectedRows>().rows().size();
}
return -1;
}
static LoD GetLoD(const Scope& scope, const std::string& name) { static LoD GetLoD(const Scope& scope, const std::string& name) {
Variable* var = scope.FindVar(name); Variable* var = scope.FindVar(name);
auto default_lod = LoD({{}}); auto default_lod = LoD({{}});
...@@ -153,6 +166,10 @@ std::string OperatorBase::DebugStringEx(const Scope* scope) const { ...@@ -153,6 +166,10 @@ std::string OperatorBase::DebugStringEx(const Scope* scope) const {
for (size_t i = 0; i < input.second.size(); ++i) { for (size_t i = 0; i < input.second.size(); ++i) {
ss << input.second[i]; ss << input.second[i];
if (scope) { if (scope) {
int row_size = GetRowSize(*scope, input.second[i]);
if (row_size >= 0) {
ss << "[row_size=" << row_size << "]";
}
ss << "[" << GetDims(*scope, input.second[i], true) << "]"; ss << "[" << GetDims(*scope, input.second[i], true) << "]";
ss << "(" << GetLoD(*scope, input.second[i]) << ")"; ss << "(" << GetLoD(*scope, input.second[i]) << ")";
} }
...@@ -173,6 +190,10 @@ std::string OperatorBase::DebugStringEx(const Scope* scope) const { ...@@ -173,6 +190,10 @@ std::string OperatorBase::DebugStringEx(const Scope* scope) const {
for (size_t i = 0; i < output.second.size(); ++i) { for (size_t i = 0; i < output.second.size(); ++i) {
ss << output.second[i]; ss << output.second[i];
if (scope) { if (scope) {
int row_size = GetRowSize(*scope, output.second[i]);
if (row_size >= 0) {
ss << "[row_size=" << row_size << "]";
}
ss << "[" << GetDims(*scope, output.second[i], true) << "]"; ss << "[" << GetDims(*scope, output.second[i], true) << "]";
ss << "(" << GetLoD(*scope, output.second[i]) << ")"; ss << "(" << GetLoD(*scope, output.second[i]) << ")";
} }
......
paddle/fluid/operators/batch_norm_mkldnn_op.cc
浏览文件 @ fd9b650d
...@@ -19,10 +19,17 @@ limitations under the License. */ ...@@ -19,10 +19,17 @@ limitations under the License. */
namespace paddle { namespace paddle {
namespace operators { namespace operators {
using Tensor = framework::Tensor; using batch_norm_bwd = mkldnn::batch_normalization_backward;
using batch_norm_fwd = mkldnn::batch_normalization_forward;
using framework::DataLayout;
using framework::Tensor;
using mkldnn::memory;
using mkldnn::primitive;
using mkldnn::reorder;
using mkldnn::stream;
using paddle::platform::MKLDNNDeviceContext; using paddle::platform::MKLDNNDeviceContext;
using paddle::platform::MKLDNNMemDesc; using paddle::platform::MKLDNNMemDesc;
using mkldnn::memory; using platform::to_void_cast;
template <typename T> template <typename T>
using EigenArrayMap = using EigenArrayMap =
...@@ -64,21 +71,12 @@ void run_batch_norm_op(Args &&... args) { ...@@ -64,21 +71,12 @@ void run_batch_norm_op(Args &&... args) {
mkldnn::stream(mkldnn::stream::kind::eager).submit(pipeline).wait(); mkldnn::stream(mkldnn::stream::kind::eager).submit(pipeline).wait();
} }
template <typename T>
inline void *cast_const_to_void(const T *t) {
return static_cast<void *>(const_cast<T *>(t));
}
} // namespace } // namespace
template <typename T> template <typename T>
class BatchNormMKLDNNOpKernel : public paddle::framework::OpKernel<T> { class BatchNormMKLDNNOpKernel : public paddle::framework::OpKernel<T> {
public: public:
void Compute(const framework::ExecutionContext &ctx) const override { void Compute(const framework::ExecutionContext &ctx) const override {
auto data_layout_str = ctx.Attr<std::string>("data_layout");
auto data_layout = framework::StringToDataLayout(data_layout_str);
PADDLE_ENFORCE(data_layout == framework::DataLayout::kNCHW,
"MKLDNN batch normalization handles only NCHW data layout");
const float epsilon = ctx.Attr<float>("epsilon"); const float epsilon = ctx.Attr<float>("epsilon");
const float momentum = ctx.Attr<float>("momentum"); const float momentum = ctx.Attr<float>("momentum");
const bool is_test = ctx.Attr<bool>("is_test"); const bool is_test = ctx.Attr<bool>("is_test");
...@@ -99,41 +97,53 @@ class BatchNormMKLDNNOpKernel : public paddle::framework::OpKernel<T> { ...@@ -99,41 +97,53 @@ class BatchNormMKLDNNOpKernel : public paddle::framework::OpKernel<T> {
const auto *scale = ctx.Input<Tensor>("Scale"); const auto *scale = ctx.Input<Tensor>("Scale");
const auto *shift = ctx.Input<Tensor>("Bias"); const auto *shift = ctx.Input<Tensor>("Bias");
y->mutable_data<T>(ctx.GetPlace()); PADDLE_ENFORCE(x->layout() == DataLayout::kMKLDNN &&
mean_out->mutable_data<T>(ctx.GetPlace()); x->format() != memory::format::format_undef,
variance_out->mutable_data<T>(ctx.GetPlace()); "Wrong layout/format set for Input x tensor");
const T *x_data = x->data<T>();
const T *mean_data = mean->data<T>();
const T *variance_data = variance->data<T>();
T *y_data = y->mutable_data<T>(ctx.GetPlace());
T *mean_out_data = mean_out->mutable_data<T>(ctx.GetPlace());
T *variance_out_data = variance_out->mutable_data<T>(ctx.GetPlace());
T *batch_mean_data = nullptr;
T *batch_variance_data = nullptr;
if (!is_test) { if (!is_test) {
batch_mean->mutable_data<T>(ctx.GetPlace()); batch_mean_data = batch_mean->mutable_data<T>(ctx.GetPlace());
batch_variance->mutable_data<T>(ctx.GetPlace()); batch_variance_data = batch_variance->mutable_data<T>(ctx.GetPlace());
} }
auto propagation = is_test == true ? mkldnn::prop_kind::forward_scoring auto propagation = is_test == true ? mkldnn::prop_kind::forward_scoring
: mkldnn::prop_kind::forward_training; : mkldnn::prop_kind::forward_training;
auto dims = paddle::framework::vectorize2int(x->dims()); auto src_tz = paddle::framework::vectorize2int(x->dims());
auto scale_tz = paddle::framework::vectorize2int(scale->dims());
auto src_md = PADDLE_ENFORCE(scale_tz.size() == 1, "Dims of scale tensor is NOT 1");
MKLDNNMemDesc(dims, memory::data_type::f32, memory::format::nchw); const unsigned int ic = scale_tz[0];
auto dst_md =
MKLDNNMemDesc(dims, memory::data_type::f32, memory::format::nchw);
auto src_pd = mkldnn::memory::primitive_desc{src_md, mkldnn_engine};
auto dst_pd = mkldnn::memory::primitive_desc{dst_md, mkldnn_engine};
auto src = mkldnn::memory{src_pd, cast_const_to_void(x->data<T>())};
auto dst = mkldnn::memory{dst_pd, y->data<T>()};
unsigned flags = mkldnn::use_scale_shift; unsigned flags = mkldnn::use_scale_shift;
if (is_test) flags |= mkldnn::use_global_stats; if (is_test) flags |= mkldnn::use_global_stats;
// create mkldnn memory from input x tensor
auto src_memory =
memory({{{src_tz}, memory::data_type::f32, x->format()}, mkldnn_engine},
to_void_cast(x_data));
// create primitive descriptor for batch norm forward
using bn_fwd_types = bn_type_traits<mkldnn::batch_normalization_forward>; using bn_fwd_types = bn_type_traits<mkldnn::batch_normalization_forward>;
auto batch_norm_fwd_desc = auto batch_norm_fwd_desc = bn_fwd_types::op_desc{
bn_fwd_types::op_desc{propagation, src_md, epsilon, flags}; propagation, src_memory.get_primitive_desc().desc(), epsilon, flags};
auto batch_norm_fwd_pd = std::shared_ptr<batch_norm_fwd::primitive_desc> batch_norm_fwd_pd =
bn_fwd_types::op_prim{batch_norm_fwd_desc, mkldnn_engine}; std::shared_ptr<batch_norm_fwd::primitive_desc>(
new batch_norm_fwd::primitive_desc(batch_norm_fwd_desc,
mkldnn_engine));
const unsigned int ic = dims[1]; // Save the pd to be used in backward pass
const std::string key = ctx.op().Output("SavedMean");
const std::string key_batch_norm_fwd_pd = key + "@bn_fwd_pd";
dev_ctx.SetBlob(key_batch_norm_fwd_pd, batch_norm_fwd_pd);
// MKLDNN requires a single piece of memory for scale and shift/bias data // MKLDNN requires a single piece of memory for scale and shift/bias data
const size_t scaleshift_size = 2 * ic; const size_t scaleshift_size = 2 * ic;
...@@ -143,73 +153,58 @@ class BatchNormMKLDNNOpKernel : public paddle::framework::OpKernel<T> { ...@@ -143,73 +153,58 @@ class BatchNormMKLDNNOpKernel : public paddle::framework::OpKernel<T> {
copy_to_weights(scale->data<T>(), scale->data<T>() + ic, shift->data<T>(), copy_to_weights(scale->data<T>(), scale->data<T>() + ic, shift->data<T>(),
shift->data<T>() + ic, &scaleshift_data); shift->data<T>() + ic, &scaleshift_data);
auto scaleshift_memory = mkldnn::memory{ // crate mkldnn memory for weights(scale/shift)
batch_norm_fwd_pd.weights_primitive_desc(), scaleshift_data.data()}; auto scaleshift_memory = memory(batch_norm_fwd_pd->weights_primitive_desc(),
scaleshift_data.data());
if (is_test) { // create mkldnn memory for output y tensor
auto mean_memory = mkldnn::memory{batch_norm_fwd_pd.mean_primitive_desc(), auto dst_memory = memory(batch_norm_fwd_pd->dst_primitive_desc(), y_data);
cast_const_to_void(mean->data<T>())};
if (is_test) {
// create mkldnn memory for stats (as input)
auto mean_memory = memory(batch_norm_fwd_pd->mean_primitive_desc(),
to_void_cast(mean_data));
auto variance_memory = auto variance_memory =
mkldnn::memory{batch_norm_fwd_pd.variance_primitive_desc(), memory(batch_norm_fwd_pd->variance_primitive_desc(),
cast_const_to_void(variance->data<T>())}; to_void_cast(variance_data));
run_batch_norm_op<typename bn_fwd_types::op_type>( run_batch_norm_op<typename bn_fwd_types::op_type>(
batch_norm_fwd_pd, src, (const mkldnn::primitive::at &)mean_memory, *batch_norm_fwd_pd, src_memory,
(const mkldnn::primitive::at &)mean_memory,
(const mkldnn::primitive::at &)variance_memory, scaleshift_memory, (const mkldnn::primitive::at &)variance_memory, scaleshift_memory,
dst); dst_memory);
} else { } else {
// create mkldnn memory for stats (as output)
auto mean_memory = auto mean_memory =
mkldnn::memory{batch_norm_fwd_pd.mean_primitive_desc(), memory(batch_norm_fwd_pd->mean_primitive_desc(), batch_mean_data);
cast_const_to_void(batch_mean->data<T>())}; auto variance_memory = memory(
batch_norm_fwd_pd->variance_primitive_desc(), batch_variance_data);
auto variance_memory =
mkldnn::memory{batch_norm_fwd_pd.variance_primitive_desc(),
cast_const_to_void(batch_variance->data<T>())};
run_batch_norm_op<bn_fwd_types::op_type>(batch_norm_fwd_pd, src, run_batch_norm_op<bn_fwd_types::op_type>(*batch_norm_fwd_pd, src_memory,
scaleshift_memory, dst, scaleshift_memory, dst_memory,
mean_memory, variance_memory); mean_memory, variance_memory);
} }
if (!is_test) { if (!is_test) {
const unsigned int in = dims[0]; // mkldnn only compute stats for current batch
const unsigned int sample_size = x->numel() / in / ic; // so we need compute momentum stats via Eigen lib
EigenVectorArrayMap<T> batch_mean_e(batch_mean_data, ic);
// saved_xx is use just in this batch of data EigenVectorArrayMap<T> batch_variance_e(batch_variance_data, ic);
EigenVectorArrayMap<T> saved_mean_e( ConstEigenVectorArrayMap<T> mean_e(mean_data, ic);
batch_mean->mutable_data<T>(ctx.GetPlace()), ic); ConstEigenVectorArrayMap<T> variance_e{variance_data, ic};
EigenVectorArrayMap<T> saved_variance_e(
batch_variance->mutable_data<T>(ctx.GetPlace()), ic);
saved_mean_e.setZero();
saved_variance_e.setZero();
const unsigned int x_arr_size = in * ic;
ConstEigenArrayMap<T> x_arr(x->data<T>(), sample_size, x_arr_size);
for (unsigned int nc = 0; nc < x_arr_size; ++nc) {
saved_mean_e(nc % ic) += x_arr.col(nc).sum();
}
saved_mean_e /= in * sample_size;
for (unsigned int nc = 0; nc < x_arr_size; ++nc) {
saved_variance_e(nc % ic) +=
(x_arr.col(nc) - saved_mean_e(nc % ic)).matrix().squaredNorm();
}
saved_variance_e /= in * sample_size;
ConstEigenVectorArrayMap<T> mean_arr{mean->data<T>(), ic};
ConstEigenVectorArrayMap<T> variance_arr{variance->data<T>(), ic};
EigenVectorArrayMap<T> running_mean_arr( EigenVectorArrayMap<T> running_mean_e(mean_out_data, ic);
mean_out->mutable_data<T>(ctx.GetPlace()), ic); EigenVectorArrayMap<T> running_variance_e(variance_out_data, ic);
EigenVectorArrayMap<T> running_var_arr(
variance_out->mutable_data<T>(ctx.GetPlace()), ic);
auto one_minus_momentum = 1. - momentum; auto one_minus_momentum = 1. - momentum;
running_mean_arr = running_mean_e = mean_e * momentum + batch_mean_e * one_minus_momentum;
mean_arr * momentum + saved_mean_e * one_minus_momentum; running_variance_e =
running_var_arr = variance_e * momentum + batch_variance_e * one_minus_momentum;
variance_arr * momentum + saved_variance_e * one_minus_momentum;
} }
y->set_layout(DataLayout::kMKLDNN);
y->set_format(
(memory::format)dst_memory.get_primitive_desc().desc().data.format);
} }
}; };
...@@ -217,11 +212,6 @@ template <typename T> ...@@ -217,11 +212,6 @@ template <typename T>
class BatchNormMKLDNNGradOpKernel : public paddle::framework::OpKernel<T> { class BatchNormMKLDNNGradOpKernel : public paddle::framework::OpKernel<T> {
public: public:
void Compute(const paddle::framework::ExecutionContext &ctx) const override { void Compute(const paddle::framework::ExecutionContext &ctx) const override {
auto data_layout_str = ctx.Attr<std::string>("data_layout");
auto data_layout = framework::StringToDataLayout(data_layout_str);
PADDLE_ENFORCE(data_layout == framework::DataLayout::kNCHW,
"MKLDNN batch normalization handles only NCHW data layout");
auto &dev_ctx = ctx.template device_context<MKLDNNDeviceContext>(); auto &dev_ctx = ctx.template device_context<MKLDNNDeviceContext>();
auto mkldnn_engine = dev_ctx.GetEngine(); auto mkldnn_engine = dev_ctx.GetEngine();
...@@ -238,88 +228,132 @@ class BatchNormMKLDNNGradOpKernel : public paddle::framework::OpKernel<T> { ...@@ -238,88 +228,132 @@ class BatchNormMKLDNNGradOpKernel : public paddle::framework::OpKernel<T> {
auto *diff_scale = ctx.Output<Tensor>(framework::GradVarName("Scale")); auto *diff_scale = ctx.Output<Tensor>(framework::GradVarName("Scale"));
auto *diff_shift = ctx.Output<Tensor>(framework::GradVarName("Bias")); auto *diff_shift = ctx.Output<Tensor>(framework::GradVarName("Bias"));
diff_x->mutable_data<T>(ctx.GetPlace()); PADDLE_ENFORCE(diff_y->layout() == DataLayout::kMKLDNN &&
diff_scale->mutable_data<T>(ctx.GetPlace()); diff_y->format() != memory::format::format_undef,
diff_shift->mutable_data<T>(ctx.GetPlace()); "Wrong layout/format set for Input diff_y tensor");
const T *x_data = x->data<T>();
const T *diff_y_data = diff_y->data<T>();
const T *batch_mean_data = batch_mean->data<T>();
const T *batch_variance_data = batch_variance->data<T>();
const T *scale_data = scale->data<T>();
const T *shift_data = shift->data<T>();
T *diff_x_data = diff_x->mutable_data<T>(ctx.GetPlace());
T *diff_scale_data = diff_scale->mutable_data<T>(ctx.GetPlace());
T *diff_shift_data = diff_shift->mutable_data<T>(ctx.GetPlace());
auto src_tz = paddle::framework::vectorize2int(x->dims());
auto diff_src_tz = src_tz;
auto dst_tz = src_tz;
auto diff_dst_tz = dst_tz;
auto scale_tz = paddle::framework::vectorize2int(scale->dims());
PADDLE_ENFORCE(scale_tz.size() == 1, "Dims of scale tensor is NOT 1");
const unsigned int ic = scale_tz[0];
// Retrieve bn_fwd_pd from device context
const std::string key = ctx.op().Input("SavedMean");
const std::string key_batch_norm_fwd_pd = key + "@bn_fwd_pd";
auto batch_norm_fwd_pd =
std::static_pointer_cast<batch_norm_fwd::primitive_desc>(
dev_ctx.GetBlob(key_batch_norm_fwd_pd));
PADDLE_ENFORCE(batch_norm_fwd_pd != nullptr,
"Fail to find batch_norm_fwd_pd in device context");
auto dims = paddle::framework::vectorize2int(x->dims()); using bn_bwd_types = bn_type_traits<mkldnn::batch_normalization_backward>;
unsigned flags = mkldnn::use_scale_shift | !mkldnn::use_global_stats;
auto src_md = // create mkldnn memory from input diff_y tensor
MKLDNNMemDesc(dims, memory::data_type::f32, memory::format::nchw); auto user_diff_dst_memory =
auto dst_md = memory({{{diff_dst_tz}, memory::data_type::f32, diff_y->format()},
MKLDNNMemDesc(dims, memory::data_type::f32, memory::format::nchw); mkldnn_engine},
auto diff_src_md = to_void_cast(diff_y_data));
MKLDNNMemDesc(dims, memory::data_type::f32, memory::format::nchw);
auto diff_dst_md =
MKLDNNMemDesc(dims, memory::data_type::f32, memory::format::nchw);
using bn_bwd_types = bn_type_traits<mkldnn::batch_normalization_backward>; // create mkldnn memory from input x tensor
using bn_fwd_types = bn_type_traits<mkldnn::batch_normalization_forward>; auto src_memory =
memory({{{src_tz}, memory::data_type::f32, x->format()}, mkldnn_engine},
to_void_cast(x_data));
auto batch_norm_fwd_desc = bn_fwd_types::op_desc{ // for diff_dst, try to use same format as dst in forward pass
mkldnn::prop_kind::forward_training, src_md, epsilon, flags}; auto diff_dst_pd = batch_norm_fwd_pd.get()->dst_primitive_desc();
auto batch_norm_fwd_pd = auto diff_dst_md = diff_dst_pd.desc();
bn_fwd_types::op_prim{batch_norm_fwd_desc, mkldnn_engine};
// create primitive descriptor for batch norm backward
unsigned flags = mkldnn::use_scale_shift;
auto batch_norm_bwd_desc = bn_bwd_types::op_desc{ auto batch_norm_bwd_desc = bn_bwd_types::op_desc{
mkldnn::prop_kind::backward, diff_dst_md, dst_md, epsilon, flags}; mkldnn::prop_kind::backward, diff_dst_md,
src_memory.get_primitive_desc().desc(), epsilon, flags};
auto batch_norm_bwd_pd = bn_bwd_types::op_prim{ auto batch_norm_bwd_pd = bn_bwd_types::op_prim{
batch_norm_bwd_desc, mkldnn_engine, batch_norm_fwd_pd}; batch_norm_bwd_desc, mkldnn_engine, *batch_norm_fwd_pd};
auto src = mkldnn::memory{{src_md, mkldnn_engine}, // reorder user_diff_dst if it's not in preferred format
cast_const_to_void(x->data<T>())}; auto diff_dst_memory = user_diff_dst_memory;
primitive reorder_diff_dst;
auto mean = mkldnn::memory{batch_norm_bwd_pd.mean_primitive_desc(), bool is_diff_dst_reordered = false;
cast_const_to_void(batch_mean->data<T>())}; if (diff_dst_pd != user_diff_dst_memory.get_primitive_desc()) {
diff_dst_memory = memory(diff_dst_pd);
auto variance = reorder_diff_dst = reorder(user_diff_dst_memory, diff_dst_memory);
mkldnn::memory{batch_norm_bwd_pd.variance_primitive_desc(), is_diff_dst_reordered = true;
cast_const_to_void(batch_variance->data<T>())}; }
auto diff_dst = mkldnn::memory{{diff_dst_md, mkldnn_engine},
cast_const_to_void(diff_y->data<T>())};
const unsigned int ic = dims[1]; // create mkldnn memory for input tensors (src/mean/variance)
auto mean_memory = memory(batch_norm_bwd_pd.mean_primitive_desc(),
to_void_cast(batch_mean_data));
auto variance_memory = memory(batch_norm_bwd_pd.variance_primitive_desc(),
to_void_cast(batch_variance_data));
// MKLDNN requires a single piece of memory for scale and shift/bias data
const size_t scaleshift_size = 2 * ic; const size_t scaleshift_size = 2 * ic;
std::vector<T> scaleshift_data; std::vector<T> scaleshift_data;
scaleshift_data.reserve(scaleshift_size); scaleshift_data.reserve(scaleshift_size);
copy_to_weights(scale->data<T>(), scale->data<T>() + ic, shift->data<T>(), copy_to_weights(scale_data, scale_data + ic, shift_data, shift_data + ic,
shift->data<T>() + ic, &scaleshift_data); &scaleshift_data);
auto scaleshift_memory = mkldnn::memory{ // create mkldnn memory for input tensors (scale/shift)
batch_norm_bwd_pd.weights_primitive_desc(), scaleshift_data.data()}; auto scaleshift_memory = memory(batch_norm_bwd_pd.weights_primitive_desc(),
scaleshift_data.data());
// create mkldnn memory for output diff weights (combined scale/shift)
std::vector<T> diff_scaleshift_data; std::vector<T> diff_scaleshift_data;
diff_scaleshift_data.reserve(scaleshift_size); diff_scaleshift_data.reserve(scaleshift_size);
copy_to_weights(diff_scale->data<T>(), diff_scale->data<T>() + ic,
diff_shift->data<T>(), diff_shift->data<T>() + ic,
&diff_scaleshift_data);
auto diff_scaleshift_memory = auto diff_scaleshift_memory =
mkldnn::memory{batch_norm_bwd_pd.diff_weights_primitive_desc(), memory(batch_norm_bwd_pd.diff_weights_primitive_desc(),
diff_scaleshift_data.data()}; diff_scaleshift_data.data());
auto diff_src = mkldnn::memory{{diff_src_md, mkldnn_engine}, // here assume diff_src is in the same format of src
static_cast<void *>(diff_x->data<T>())}; auto diff_src_memory = memory(src_memory.get_primitive_desc(), diff_x_data);
run_batch_norm_op<bn_bwd_types::op_type>( // finally create batch_norm backward primitive
batch_norm_bwd_pd, src, mean, variance, diff_dst, scaleshift_memory, auto batch_norm_bwd_prim =
diff_src, diff_scaleshift_memory); batch_norm_bwd(batch_norm_bwd_pd, src_memory, mean_memory,
variance_memory, diff_dst_memory, scaleshift_memory,
diff_src_memory, diff_scaleshift_memory);
// execute optional reorder and batch_norm backward primitive
std::vector<primitive> pipeline;
if (is_diff_dst_reordered) pipeline.push_back(reorder_diff_dst);
pipeline.push_back(batch_norm_bwd_prim);
stream(stream::kind::eager).submit(pipeline).wait();
// copy back diff sacle/shift to output tensors (diff scale/shift)
diff_scaleshift_data.resize(scaleshift_size);
auto it = std::begin(diff_scaleshift_data); auto it = std::begin(diff_scaleshift_data);
std::copy(it, std::next(it, ic), diff_scale->data<T>()); std::copy(it, std::next(it, ic), diff_scale_data);
std::copy(std::next(it, ic), std::end(diff_scaleshift_data), std::copy(std::next(it, ic), std::end(diff_scaleshift_data),
diff_shift->data<T>()); diff_shift_data);
// set layout/format of output tensors
diff_x->set_layout(DataLayout::kMKLDNN);
diff_x->set_format((memory::format)diff_src_memory.get_primitive_desc()
.desc()
.data.format);
} }
}; };
} // namespace operators } // namespace operators
} // namespace paddle } // namespace paddle
namespace ops = paddle::operators; namespace ops = paddle::operators;
REGISTER_OP_KERNEL(batch_norm, MKLDNN, paddle::platform::CPUPlace, REGISTER_OP_KERNEL(batch_norm, MKLDNN, ::paddle::platform::CPUPlace,
ops::BatchNormMKLDNNOpKernel<float>); ops::BatchNormMKLDNNOpKernel<float>);
REGISTER_OP_KERNEL(batch_norm_grad, MKLDNN, paddle::platform::CPUPlace, REGISTER_OP_KERNEL(batch_norm_grad, MKLDNN, ::paddle::platform::CPUPlace,
ops::BatchNormMKLDNNGradOpKernel<float>); ops::BatchNormMKLDNNGradOpKernel<float>);
paddle/fluid/operators/batch_norm_op.cc
浏览文件 @ fd9b650d
...@@ -110,19 +110,19 @@ class BatchNormOp : public framework::OperatorWithKernel { ...@@ -110,19 +110,19 @@ class BatchNormOp : public framework::OperatorWithKernel {
ctx.Input<Tensor>("Variance")->type()), ctx.Input<Tensor>("Variance")->type()),
"Variance input should be of float type"); "Variance input should be of float type");
framework::LibraryType library_{framework::LibraryType::kPlain};
// TODO(pzelazko-intel): enable MKLDNN layout when it's ready // TODO(pzelazko-intel): enable MKLDNN layout when it's ready
framework::LibraryType library = framework::LibraryType::kPlain;
framework::DataLayout layout = framework::DataLayout::kAnyLayout; framework::DataLayout layout = framework::DataLayout::kAnyLayout;
#ifdef PADDLE_WITH_MKLDNN #ifdef PADDLE_WITH_MKLDNN
if (library_ == framework::LibraryType::kPlain && if (library == framework::LibraryType::kPlain &&
platform::CanMKLDNNBeUsed(ctx)) { platform::CanMKLDNNBeUsed(ctx)) {
library_ = framework::LibraryType::kMKLDNN; library = framework::LibraryType::kMKLDNN;
layout = framework::DataLayout::kMKLDNN; layout = framework::DataLayout::kMKLDNN;
} }
#endif #endif
return framework::OpKernelType(input_data_type, ctx.GetPlace(), layout, return framework::OpKernelType(input_data_type, ctx.GetPlace(), layout,
library_); library);
} }
}; };
...@@ -370,19 +370,21 @@ class BatchNormGradOp : public framework::OperatorWithKernel { ...@@ -370,19 +370,21 @@ class BatchNormGradOp : public framework::OperatorWithKernel {
PADDLE_THROW("can't find Y@GRAD"); PADDLE_THROW("can't find Y@GRAD");
} }
framework::LibraryType library_{framework::LibraryType::kPlain};
// TODO(pzelazko-intel): enable MKLDNN layout when it's ready // TODO(pzelazko-intel): enable MKLDNN layout when it's ready
framework::DataLayout layout_ = framework::DataLayout::kAnyLayout; framework::LibraryType library = framework::LibraryType::kPlain;
framework::DataLayout layout = framework::DataLayout::kAnyLayout;
#ifdef PADDLE_WITH_MKLDNN #ifdef PADDLE_WITH_MKLDNN
if (library_ == framework::LibraryType::kPlain && if (library == framework::LibraryType::kPlain &&
platform::CanMKLDNNBeUsed(ctx)) { platform::CanMKLDNNBeUsed(ctx)) {
library_ = framework::LibraryType::kMKLDNN; library = framework::LibraryType::kMKLDNN;
layout_ = framework::DataLayout::kMKLDNN; layout = framework::DataLayout::kMKLDNN;
} }
#endif #endif
return framework::OpKernelType( return framework::OpKernelType(
framework::ToDataType(ctx.Input<Tensor>("X")->type()), ctx.GetPlace(), framework::ToDataType(ctx.Input<Tensor>("X")->type()), ctx.GetPlace(),
layout_, library_); layout, library);
} }
}; };
......
paddle/fluid/operators/conv_mkldnn_op.cc
浏览文件 @ fd9b650d
...@@ -18,6 +18,17 @@ ...@@ -18,6 +18,17 @@
namespace paddle { namespace paddle {
namespace operators { namespace operators {
using conv_bwd_data = mkldnn::convolution_backward_data;
using conv_bwd_weights = mkldnn::convolution_backward_weights;
using conv_fwd = mkldnn::convolution_forward;
using framework::DataLayout;
using mkldnn::memory;
using mkldnn::primitive;
using mkldnn::reorder;
using mkldnn::stream;
using platform::to_void_cast;
using platform::GetMKLDNNFormat;
template <typename T> template <typename T>
class ConvMKLDNNOpKernel : public paddle::framework::OpKernel<T> { class ConvMKLDNNOpKernel : public paddle::framework::OpKernel<T> {
public: public:
...@@ -25,6 +36,10 @@ class ConvMKLDNNOpKernel : public paddle::framework::OpKernel<T> { ...@@ -25,6 +36,10 @@ class ConvMKLDNNOpKernel : public paddle::framework::OpKernel<T> {
PADDLE_ENFORCE(paddle::platform::is_cpu_place(ctx.GetPlace()), PADDLE_ENFORCE(paddle::platform::is_cpu_place(ctx.GetPlace()),
"It must use CPUPlace."); "It must use CPUPlace.");
// Get unique name for index
const std::string key = ctx.op().Output("Output");
const std::string key_conv_pd = key + "@conv_pd";
auto& dev_ctx = auto& dev_ctx =
ctx.template device_context<paddle::platform::MKLDNNDeviceContext>(); ctx.template device_context<paddle::platform::MKLDNNDeviceContext>();
const auto& mkldnn_engine = dev_ctx.GetEngine(); const auto& mkldnn_engine = dev_ctx.GetEngine();
...@@ -33,10 +48,12 @@ class ConvMKLDNNOpKernel : public paddle::framework::OpKernel<T> { ...@@ -33,10 +48,12 @@ class ConvMKLDNNOpKernel : public paddle::framework::OpKernel<T> {
auto* filter = ctx.Input<Tensor>("Filter"); auto* filter = ctx.Input<Tensor>("Filter");
auto* output = ctx.Output<Tensor>("Output"); auto* output = ctx.Output<Tensor>("Output");
// Get an unique name from "argument" name of "Output" variable PADDLE_ENFORCE(input->layout() == DataLayout::kMKLDNN &&
// This name will be used as key when saving info into device context input->format() != memory::format::format_undef,
const std::string key = ctx.op().Output("Output"); "Wrong layout/format set for Input tensor");
const std::string key_conv_pd = key + "@conv_pd"; PADDLE_ENFORCE(filter->layout() == DataLayout::kMKLDNN &&
filter->format() != memory::format::format_undef,
"Wrong layout/format set for Filter tensor");
std::vector<int> strides = ctx.Attr<std::vector<int>>("strides"); std::vector<int> strides = ctx.Attr<std::vector<int>>("strides");
std::vector<int> paddings = ctx.Attr<std::vector<int>>("paddings"); std::vector<int> paddings = ctx.Attr<std::vector<int>>("paddings");
...@@ -63,60 +80,86 @@ class ConvMKLDNNOpKernel : public paddle::framework::OpKernel<T> { ...@@ -63,60 +80,86 @@ class ConvMKLDNNOpKernel : public paddle::framework::OpKernel<T> {
paddle::framework::vectorize2int(filter->dims()); paddle::framework::vectorize2int(filter->dims());
std::vector<int> dst_tz = paddle::framework::vectorize2int(output->dims()); std::vector<int> dst_tz = paddle::framework::vectorize2int(output->dims());
// TODO(pzelazko-intel): support more formats // create mkldnn memory from input tensors (data/weights)
auto src_md = platform::MKLDNNMemDesc( auto user_src_memory = memory(
src_tz, mkldnn::memory::data_type::f32, mkldnn::memory::format::nchw); {{{src_tz}, memory::data_type::f32, input->format()}, mkldnn_engine},
auto weights_md = to_void_cast(input_data));
platform::MKLDNNMemDesc(weights_tz, mkldnn::memory::data_type::f32, auto user_weights_memory =
mkldnn::memory::format::oihw); memory({{{weights_tz}, memory::data_type::f32, filter->format()},
auto dst_md = platform::MKLDNNMemDesc( mkldnn_engine},
dst_tz, mkldnn::memory::data_type::f32, mkldnn::memory::format::nchw); to_void_cast(filter_data));
auto src_memory = /* create memory descriptor for convolution without specified format
mkldnn::memory({src_md, mkldnn_engine}, * ('any') which lets a primitive (convolution in this case) choose
reinterpret_cast<void*>(const_cast<T*>(input_data))); * the memory format preferred for best performance
auto weights_memory = */
mkldnn::memory({weights_md, mkldnn_engine}, auto src_md = platform::MKLDNNMemDesc(src_tz, memory::data_type::f32,
reinterpret_cast<void*>(const_cast<T*>(filter_data))); memory::format::any);
auto dst_memory = mkldnn::memory({dst_md, mkldnn_engine}, output_data); auto weights_md = platform::MKLDNNMemDesc(
weights_tz, memory::data_type::f32, memory::format::any);
std::shared_ptr<mkldnn::convolution_forward::primitive_desc> conv_pd = auto dst_md = platform::MKLDNNMemDesc(dst_tz, memory::data_type::f32,
ConvFwdPrimitiveDesc(src_md, weights_md, dst_md, strides, paddings, memory::format::any);
mkldnn_engine);
// create a conv primitive descriptor and save it for usage in backward
// save conv_pd into global device context to be referred in backward path std::shared_ptr<conv_fwd::primitive_desc> conv_pd = ConvFwdPrimitiveDesc(
dev_ctx.SetBlob(key_conv_pd, conv_pd); src_md, weights_md, dst_md, strides, paddings, mkldnn_engine);
// create reorder primitive if the input format is not the preferred one
auto src_memory = user_src_memory;
primitive reorder_src;
bool is_src_reordered = false;
if (memory::primitive_desc(conv_pd->src_primitive_desc()) !=
user_src_memory.get_primitive_desc()) {
src_memory = memory(conv_pd->src_primitive_desc());
reorder_src = reorder(user_src_memory, src_memory);
is_src_reordered = true;
}
auto weights_memory = user_weights_memory;
primitive reorder_weights;
bool is_weights_reordered = false;
if (memory::primitive_desc(conv_pd->weights_primitive_desc()) !=
user_weights_memory.get_primitive_desc()) {
weights_memory = memory(conv_pd->weights_primitive_desc());
reorder_weights = reorder(user_weights_memory, weights_memory);
is_weights_reordered = true;
}
// create memory primitive for conv dst
auto dst_memory = memory(conv_pd->dst_primitive_desc(), output_data);
// create convolution op primitive // create convolution op primitive
auto conv_prim = mkldnn::convolution_forward(*conv_pd, src_memory, auto conv_prim = conv_fwd(*conv_pd, src_memory, weights_memory, dst_memory);
weights_memory, dst_memory);
// push primitive to stream and wait until it's executed // push primitive to stream and wait until it's executed
std::vector<mkldnn::primitive> pipeline{conv_prim}; std::vector<primitive> pipeline;
mkldnn::stream(mkldnn::stream::kind::eager).submit(pipeline).wait(); if (is_src_reordered) pipeline.push_back(reorder_src);
if (is_weights_reordered) pipeline.push_back(reorder_weights);
pipeline.push_back(conv_prim);
stream(stream::kind::eager).submit(pipeline).wait();
// Save conv_pd/src_memory/weights_memory for backward pass
dev_ctx.SetBlob(key_conv_pd, conv_pd);
output->set_layout(DataLayout::kMKLDNN);
output->set_format(GetMKLDNNFormat(dst_memory));
} }
private: private:
std::unique_ptr<mkldnn::convolution_forward::primitive_desc> std::unique_ptr<conv_fwd::primitive_desc> ConvFwdPrimitiveDesc(
ConvFwdPrimitiveDesc(const mkldnn::memory::desc& src, const memory::desc& src, const memory::desc& weights,
const mkldnn::memory::desc& weights, const memory::desc& dst, const std::vector<int>& strides,
const mkldnn::memory::desc& dst, const std::vector<int>& paddings, const mkldnn::engine& engine) const {
const std::vector<int>& strides, memory::dims stride_dims = {strides[0], strides[1]};
const std::vector<int>& paddings, memory::dims padding_dims = {paddings[0], paddings[1]};
const mkldnn::engine& engine) const {
mkldnn::memory::dims stride_dims = {strides[0], strides[1]}; auto conv_desc =
mkldnn::memory::dims padding_dims = {paddings[0], paddings[1]}; conv_fwd::desc(mkldnn::prop_kind::forward, mkldnn::convolution_direct,
src, weights, dst, stride_dims, padding_dims,
auto conv_desc = mkldnn::convolution_forward::desc( padding_dims, mkldnn::padding_kind::zero);
mkldnn::prop_kind::forward, mkldnn::convolution_direct, src, weights,
dst, stride_dims, padding_dims, padding_dims, auto p_conv_pd = new conv_fwd::primitive_desc(conv_desc, engine);
mkldnn::padding_kind::zero);
return std::unique_ptr<conv_fwd::primitive_desc>(p_conv_pd);
auto p_conv_pd =
new mkldnn::convolution_forward::primitive_desc(conv_desc, engine);
return std::unique_ptr<mkldnn::convolution_forward::primitive_desc>(
p_conv_pd);
} }
}; };
...@@ -139,6 +182,19 @@ class ConvMKLDNNGradOpKernel : public paddle::framework::OpKernel<T> { ...@@ -139,6 +182,19 @@ class ConvMKLDNNGradOpKernel : public paddle::framework::OpKernel<T> {
Tensor* input_grad = ctx.Output<Tensor>(framework::GradVarName("Input")); Tensor* input_grad = ctx.Output<Tensor>(framework::GradVarName("Input"));
Tensor* filter_grad = ctx.Output<Tensor>(framework::GradVarName("Filter")); Tensor* filter_grad = ctx.Output<Tensor>(framework::GradVarName("Filter"));
PADDLE_ENFORCE(input->layout() == DataLayout::kMKLDNN &&
input->format() != memory::format::format_undef,
"Wrong layout/format set for Input tensor");
PADDLE_ENFORCE(filter->layout() == DataLayout::kMKLDNN &&
filter->format() != memory::format::format_undef,
"Wrong layout/format set for Filter tensor");
PADDLE_ENFORCE(output->layout() == DataLayout::kMKLDNN &&
output->format() != memory::format::format_undef,
"Wrong layout/format set for Output tensor");
PADDLE_ENFORCE(output_grad->layout() == DataLayout::kMKLDNN &&
output_grad->format() != memory::format::format_undef,
"Wrong layout/format set for output_grad tensor");
if (!input_grad && !filter_grad) return; if (!input_grad && !filter_grad) return;
// Get an unique name from "argument" name of "Output" variable // Get an unique name from "argument" name of "Output" variable
...@@ -167,108 +223,147 @@ class ConvMKLDNNGradOpKernel : public paddle::framework::OpKernel<T> { ...@@ -167,108 +223,147 @@ class ConvMKLDNNGradOpKernel : public paddle::framework::OpKernel<T> {
paddle::framework::vectorize2int(filter->dims()); paddle::framework::vectorize2int(filter->dims());
std::vector<int> dst_tz = paddle::framework::vectorize2int(output->dims()); std::vector<int> dst_tz = paddle::framework::vectorize2int(output->dims());
// TODO(pzelazko-intel): support more formats // create mkldnn memory from input tensors (input/weights/output_grad)
auto src_md = platform::MKLDNNMemDesc( auto user_src_memory = memory(
src_tz, mkldnn::memory::data_type::f32, mkldnn::memory::format::nchw); {{{src_tz}, memory::data_type::f32, input->format()}, mkldnn_engine},
auto diff_src_md = platform::MKLDNNMemDesc( to_void_cast(input_data));
src_tz, mkldnn::memory::data_type::f32, mkldnn::memory::format::nchw); auto user_weights_memory =
auto weights_md = memory({{{weights_tz}, memory::data_type::f32, filter->format()},
platform::MKLDNNMemDesc(weights_tz, mkldnn::memory::data_type::f32, mkldnn_engine},
mkldnn::memory::format::oihw); to_void_cast(filter_data));
auto diff_weights_md = auto user_diff_dst_memory =
platform::MKLDNNMemDesc(weights_tz, mkldnn::memory::data_type::f32, memory({{{dst_tz}, memory::data_type::f32, output_grad->format()},
mkldnn::memory::format::oihw); mkldnn_engine},
auto diff_dst_md = platform::MKLDNNMemDesc( to_void_cast(output_grad_data));
dst_tz, mkldnn::memory::data_type::f32, mkldnn::memory::format::nchw);
/* create memory descriptor for conv backward without specified format
// create memory * ('any') which lets a primitive (conv backward in this case) choose
auto diff_dst_memory = mkldnn::memory( * the memory format preferred for best performance
{diff_weights_md, mkldnn_engine}, */
reinterpret_cast<void*>(const_cast<T*>(output_grad_data))); auto src_md = platform::MKLDNNMemDesc(src_tz, memory::data_type::f32,
memory::format::any);
auto diff_src_md = platform::MKLDNNMemDesc(src_tz, memory::data_type::f32,
memory::format::any);
auto weights_md = platform::MKLDNNMemDesc(
weights_tz, memory::data_type::f32, memory::format::any);
auto diff_weights_md = platform::MKLDNNMemDesc(
weights_tz, memory::data_type::f32, memory::format::any);
auto diff_dst_md = platform::MKLDNNMemDesc(dst_tz, memory::data_type::f32,
memory::format::any);
// Retrieve conv_pd from device context // Retrieve conv_pd from device context
auto conv_pd = auto conv_pd = std::static_pointer_cast<conv_fwd::primitive_desc>(
std::static_pointer_cast<mkldnn::convolution_forward::primitive_desc>(
dev_ctx.GetBlob(key_conv_pd)); dev_ctx.GetBlob(key_conv_pd));
PADDLE_ENFORCE(conv_pd != nullptr, PADDLE_ENFORCE(conv_pd != nullptr,
"Fail to find conv_pd in device context"); "Fail to find conv_pd in device context");
// create backward conv primitive for weights // create backward conv primitive for weights
if (filter_grad) { if (filter_grad) {
// create primitive descriptor // create backward convolution primitive descriptor
mkldnn::convolution_backward_weights::primitive_desc conv_bwd_weights_pd = auto conv_bwd_weights_desc = conv_bwd_weights::desc(
ConvBwdWeightsPrimitiveDesc(src_md, diff_weights_md, diff_dst_md, mkldnn::convolution_direct, src_md, diff_weights_md, diff_dst_md,
strides, paddings, *conv_pd, strides, paddings, paddings, mkldnn::padding_kind::zero);
mkldnn_engine); auto conv_bwd_weights_pd = conv_bwd_weights::primitive_desc(
conv_bwd_weights_desc, mkldnn_engine, *conv_pd);
// create reorder primitive if the input format is not the preferred one
auto src_memory = user_src_memory;
primitive reorder_src;
bool is_src_reordered = false;
if (memory::primitive_desc(conv_bwd_weights_pd.src_primitive_desc()) !=
user_src_memory.get_primitive_desc()) {
src_memory = memory(conv_bwd_weights_pd.src_primitive_desc());
reorder_src = reorder(user_src_memory, src_memory);
is_src_reordered = true;
}
// create memory auto diff_dst_memory_4filter = user_diff_dst_memory;
primitive reorder_diff_dst_4filter;
bool is_diff_dst_reordered_4filter = false;
if (memory::primitive_desc(
conv_bwd_weights_pd.diff_dst_primitive_desc()) !=
user_diff_dst_memory.get_primitive_desc()) {
diff_dst_memory_4filter =
memory(conv_bwd_weights_pd.diff_dst_primitive_desc());
reorder_diff_dst_4filter =
reorder(user_diff_dst_memory, diff_dst_memory_4filter);
is_diff_dst_reordered_4filter = true;
}
// create mkldnn memory for output (i.e. diff weights)
auto diff_weights_memory = auto diff_weights_memory =
mkldnn::memory({diff_weights_md, mkldnn_engine}, memory(conv_bwd_weights_pd.diff_weights_primitive_desc(),
reinterpret_cast<void*>(filter_grad_data)); reinterpret_cast<void*>(filter_grad_data));
auto src_memory =
mkldnn::memory({src_md, mkldnn_engine},
reinterpret_cast<void*>(const_cast<T*>(input_data)));
// create backward conv primitive for weights // create backward conv primitive for weights
auto conv_bwd_weights_prim = mkldnn::convolution_backward_weights( auto conv_bwd_weights_prim =
conv_bwd_weights_pd, src_memory, diff_dst_memory, conv_bwd_weights(conv_bwd_weights_pd, src_memory,
diff_weights_memory); diff_dst_memory_4filter, diff_weights_memory);
// push primitive and execute it // push primitive and execute it
std::vector<mkldnn::primitive> pipeline{conv_bwd_weights_prim}; std::vector<primitive> pipeline;
mkldnn::stream(mkldnn::stream::kind::eager).submit(pipeline).wait(); if (is_src_reordered) pipeline.push_back(reorder_src);
if (is_diff_dst_reordered_4filter)
pipeline.push_back(reorder_diff_dst_4filter);
pipeline.push_back(conv_bwd_weights_prim);
stream(stream::kind::eager).submit(pipeline).wait();
filter_grad->set_layout(DataLayout::kMKLDNN);
filter_grad->set_format(GetMKLDNNFormat(diff_weights_memory));
} }
if (input_grad) { if (input_grad) {
// create primitive descriptor // create backward convolution primitive descriptor
mkldnn::convolution_backward_data::primitive_desc conv_bwd_data_pd = auto conv_bwd_data_desc = conv_bwd_data::desc(
ConvBwdDataPrimitiveDesc(diff_src_md, weights_md, diff_dst_md, mkldnn::convolution_direct, diff_src_md, weights_md, diff_dst_md,
strides, paddings, *conv_pd, mkldnn_engine); strides, paddings, paddings, mkldnn::padding_kind::zero);
auto conv_bwd_data_pd = conv_bwd_data::primitive_desc(
// create memory conv_bwd_data_desc, mkldnn_engine, *conv_pd);
auto diff_src_memory = mkldnn::memory(
{diff_src_md, mkldnn_engine}, // create reorder primitive if the input format is not the preferred one
reinterpret_cast<void*>(const_cast<T*>(input_grad_data))); auto weights_memory = user_weights_memory;
auto weights_memory = primitive reorder_weights;
mkldnn::memory({weights_md, mkldnn_engine}, bool is_weights_reordered = false;
reinterpret_cast<void*>(const_cast<T*>(filter_data))); if (memory::primitive_desc(conv_bwd_data_pd.weights_primitive_desc()) !=
user_weights_memory.get_primitive_desc()) {
// create backward conv primitive for data weights_memory = memory(conv_bwd_data_pd.weights_primitive_desc());
auto conv_bwd_data_prim = mkldnn::convolution_backward_data( reorder_weights = reorder(user_weights_memory, weights_memory);
conv_bwd_data_pd, diff_dst_memory, weights_memory, diff_src_memory); is_weights_reordered = true;
// push primitive to stream and wait until it's executed
std::vector<mkldnn::primitive> pipeline{conv_bwd_data_prim};
mkldnn::stream(mkldnn::stream::kind::eager).submit(pipeline).wait();
} }
} // Compute()
private: auto diff_dst_memory_4data = user_diff_dst_memory;
mkldnn::convolution_backward_weights::primitive_desc primitive reorder_diff_dst_4data;
ConvBwdWeightsPrimitiveDesc( bool is_diff_dst_reordered_4data = false;
const mkldnn::memory::desc& src, const mkldnn::memory::desc& diff_weights, if (memory::primitive_desc(conv_bwd_data_pd.diff_dst_primitive_desc()) !=
const mkldnn::memory::desc& diff_dst, const std::vector<int>& strides, user_diff_dst_memory.get_primitive_desc()) {
const std::vector<int>& paddings, diff_dst_memory_4data =
const mkldnn::convolution_forward::primitive_desc& conv_pd, memory(conv_bwd_data_pd.diff_dst_primitive_desc());
const mkldnn::engine& engine) const { reorder_diff_dst_4data =
auto conv_bwd_weights_desc = mkldnn::convolution_backward_weights::desc( reorder(user_diff_dst_memory, diff_dst_memory_4data);
mkldnn::convolution_direct, src, diff_weights, diff_dst, strides, is_diff_dst_reordered_4data = true;
paddings, paddings, mkldnn::padding_kind::zero);
return mkldnn::convolution_backward_weights::primitive_desc(
conv_bwd_weights_desc, engine, conv_pd);
} }
mkldnn::convolution_backward_data::primitive_desc ConvBwdDataPrimitiveDesc( // create mkldnn memory for output (i.e. diff src)
const mkldnn::memory::desc& diff_src, const mkldnn::memory::desc& weights, auto diff_src_memory = memory(conv_bwd_data_pd.diff_src_primitive_desc(),
const mkldnn::memory::desc& diff_dst, const std::vector<int>& strides, reinterpret_cast<void*>(input_grad_data));
const std::vector<int>& paddings,
const mkldnn::convolution_forward::primitive_desc& conv_pd, // create backward conv primitive for data
const mkldnn::engine& engine) const { auto conv_bwd_data_prim =
auto conv_bwd_data_desc = mkldnn::convolution_backward_data::desc( conv_bwd_data(conv_bwd_data_pd, diff_dst_memory_4data, weights_memory,
mkldnn::convolution_direct, diff_src, weights, diff_dst, strides, diff_src_memory);
paddings, paddings, mkldnn::padding_kind::zero);
return mkldnn::convolution_backward_data::primitive_desc(conv_bwd_data_desc, // push primitive and execute it
engine, conv_pd); std::vector<primitive> pipeline;
if (is_weights_reordered) pipeline.push_back(reorder_weights);
if (is_diff_dst_reordered_4data)
pipeline.push_back(reorder_diff_dst_4data);
pipeline.push_back(conv_bwd_data_prim);
stream(stream::kind::eager).submit(pipeline).wait();
input_grad->set_layout(DataLayout::kMKLDNN);
input_grad->set_format(GetMKLDNNFormat(diff_src_memory));
} }
} // Compute()
}; };
} // namespace operators } // namespace operators
......
paddle/fluid/operators/conv_op.cc
浏览文件 @ fd9b650d
...@@ -75,9 +75,8 @@ void ConvOp::InferShape(framework::InferShapeContext* ctx) const { ...@@ -75,9 +75,8 @@ void ConvOp::InferShape(framework::InferShapeContext* ctx) const {
framework::OpKernelType ConvOp::GetExpectedKernelType( framework::OpKernelType ConvOp::GetExpectedKernelType(
const framework::ExecutionContext& ctx) const { const framework::ExecutionContext& ctx) const {
framework::LibraryType library{framework::LibraryType::kPlain}; framework::LibraryType library{framework::LibraryType::kPlain};
std::string data_format = ctx.Attr<std::string>("data_format");
// TODO(pzelazko-intel): enable MKLDNN layout when it's ready // TODO(pzelazko-intel): enable MKLDNN layout when it's ready
std::string data_format = ctx.Attr<std::string>("data_format");
framework::DataLayout layout = framework::StringToDataLayout(data_format); framework::DataLayout layout = framework::StringToDataLayout(data_format);
#ifdef PADDLE_WITH_CUDA #ifdef PADDLE_WITH_CUDA
......
paddle/fluid/operators/gen_nccl_id_op.cc
浏览文件 @ fd9b650d
...@@ -67,6 +67,10 @@ class GenNCCLIdOp : public framework::OperatorBase { ...@@ -67,6 +67,10 @@ class GenNCCLIdOp : public framework::OperatorBase {
client->AsyncSendVar(ep, dev_ctx, *scope, NCCL_ID_VARNAME); client->AsyncSendVar(ep, dev_ctx, *scope, NCCL_ID_VARNAME);
} }
client->Wait(); client->Wait();
for (auto& ep : endpoint_list) {
client->AsyncSendBatchBarrier(ep);
}
client->Wait();
VLOG(3) << "sending completed..."; VLOG(3) << "sending completed...";
} }
......
paddle/fluid/operators/merge_ids_op.cc 0 → 100644
浏览文件 @ fd9b650d
/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License. */
#include "paddle/fluid/operators/merge_ids_op.h"
namespace paddle {
namespace operators {
class MergeIdsOpMaker : public framework::OpProtoAndCheckerMaker {
public:
void Make() override {
AddInput("Ids", "(LoDTensor) the input ids with shape{batch_num, 1}");
AddInput(
"X",
"(LoDTensors) multi input tensor with shape{batch_num, N}, N is the "
"size of embedding table")
.AsDuplicable();
AddOutput("Out", "(LoDTensor) The merged outputs of the input tensors.");
AddComment(R"DOC(
Merge multi LoDTensor's into one according to Ids's shard num.
split_ids_op -> prefetch_op -> merge_ids_op
merge_ids_op should be used after split_ids_op and prefetch_op, split_ids_op
will split input Ids into multiple tensors according to Id's shard number.
prefetch_op will send them to parameter server to prefetch embedding value
back. During split, the order of ids is disordered. In merge_ids_op we use
the original Ids to restore the order of the fetched embedding value and
also pass the lod information to the merged output.
Example:
Ids = [1,2,3,4,5,6] # 3 shared
split_ids_op ->
Id0 = [3, 6] # id % 3 == 0
Id1 = [1, 4] # id % 3 == 1
Id2 = [2, 5] # id % 3 == 2
prefetch_op ->
X0 = [[0.3 0.3] # 3
[0.6 0.6]] # 6
X1 = [[0.1 0.1] # 1
[0.4 0.4]] # 4
X2 = [[0.2 0.2] # 2
[0.5 0.5]] # 5
merge_ids_op ->
Out = [[0.1 0.1] # 1
[0.2 0.2] # 2
[0.3 0.3] # 3
[0.4 0.4] # 4
[0.5 0.5] # 5
[0.6 0.6]] # 6
)DOC");
}
};
class MergeIdsOp : public framework::OperatorWithKernel {
public:
using framework::OperatorWithKernel::OperatorWithKernel;
void InferShape(framework::InferShapeContext *ctx) const override {
PADDLE_ENFORCE(ctx->HasInput("Ids"), "MergeIdsOp must has input Ids.");
PADDLE_ENFORCE(ctx->HasInputs("X"), "MergeIdsOp must has input X.");
PADDLE_ENFORCE(ctx->HasOutput("Out"), "MergeIdsOp must has output Out.");
auto ids_var_type = ctx->GetInputsVarType("Ids").front();
auto ids_dims = ctx->GetInputDim("Ids");
if (ids_var_type == framework::proto::VarType::LOD_TENSOR) {
PADDLE_ENFORCE_EQ(ids_dims.size(), 2);
PADDLE_ENFORCE_EQ(ids_dims[1], 1);
}
auto x_var_type = ctx->GetInputsVarType("X");
for (auto &var_type : x_var_type) {
PADDLE_ENFORCE_EQ(var_type, framework::proto::VarType::LOD_TENSOR,
"input X only support lod tensors");
}
ctx->ShareLoD("Ids", "Out");
}
private:
framework::OpKernelType GetExpectedKernelType(
const framework::ExecutionContext &ctx) const override {
return framework::OpKernelType(
framework::ToDataType(
ctx.MultiInput<framework::Tensor>("X").front()->type()),
ctx.GetPlace());
}
};
class MergeIdsOpInferVarType : public framework::VarTypeInference {
public:
void operator()(const framework::OpDesc &op_desc,
framework::BlockDesc *block) const override {
auto *input_var = block->Var(op_desc.Input("Ids")[0]);
for (auto &out_var : op_desc.Output("Out")) {
block->Var(out_var)->SetType(input_var->GetType());
}
}
};
} // namespace operators
} // namespace paddle
namespace ops = paddle::operators;
REGISTER_OPERATOR(merge_ids, ops::MergeIdsOp, ops::MergeIdsOpMaker,
ops::MergeIdsOpInferVarType);
REGISTER_OP_CPU_KERNEL(
merge_ids, ops::MergeIdsOpKernel<paddle::platform::CPUPlace, float>);
paddle/fluid/operators/merge_ids_op.h 0 → 100644
浏览文件 @ fd9b650d
/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License. */
#pragma once
#include <vector>
#include "paddle/fluid/framework/op_registry.h"
#include "paddle/fluid/framework/tensor_util.h"
#include "paddle/fluid/operators/math/selected_rows_functor.h"
namespace paddle {
namespace operators {
template <typename DeviceContext, typename T>
class MergeIdsOpKernel : public framework::OpKernel<T> {
public:
void Compute(const framework::ExecutionContext &ctx) const override {
auto place = ctx.GetPlace();
if (!platform::is_cpu_place(place)) {
PADDLE_THROW("MergeIds do not support GPU kernel");
}
VLOG(3) << "run in MergeIdsOpKernel";
const auto *ids_var = ctx.InputVar("Ids");
PADDLE_ENFORCE(ids_var->IsType<framework::LoDTensor>(),
"only support to merge Ids of LoDTensor");
const auto &ids_tensor = ids_var->Get<framework::LoDTensor>();
const auto &ids_dims = ids_tensor.dims();
const int64_t *ids = ids_tensor.data<int64_t>();
auto x_tensors = ctx.MultiInput<framework::LoDTensor>("X");
auto *out = ctx.Output<framework::LoDTensor>("Out");
int batch_size = 0;
int embedding_size = 0;
for (auto &input : x_tensors) {
if (framework::product(input->dims()) != 0) {
if (embedding_size == 0) {
embedding_size = input->dims()[1];
}
PADDLE_ENFORCE_EQ(embedding_size, input->dims()[1],
"embedding size of all input should be the same");
batch_size += input->dims()[0];
}
}
PADDLE_ENFORCE_EQ(
batch_size, ids_dims[0],
"the batch size of ids and merged embedding value should be the same");
const size_t shard_num = x_tensors.size();
if (shard_num == 1) {
VLOG(3) << "only one shard, we can copy the data directly";
TensorCopy(*x_tensors[0], place, out);
} else {
std::vector<int> in_indexs(shard_num, 0);
auto *out_data = out->mutable_data<T>(
framework::make_ddim({batch_size, embedding_size}), place);
// copy data from ins[shard_num] to out.
for (int i = 0; i < ids_dims[0]; ++i) {
int64_t id = ids[i];
size_t shard_id = static_cast<size_t>(id) % shard_num;
int index = in_indexs[shard_id];
memcpy(out_data + embedding_size * i,
x_tensors[shard_id]->data<T>() + index * embedding_size,
sizeof(T) * embedding_size);
in_indexs[shard_id] += 1;
}
for (size_t i = 0; i < shard_num; ++i) {
PADDLE_ENFORCE_EQ(in_indexs[i], x_tensors[i]->dims()[0],
"after merge, all data in x_tensor should be used");
}
}
}
};
} // namespace operators
} // namespace paddle
paddle/fluid/platform/cpu_info.cc
浏览文件 @ fd9b650d
...@@ -59,9 +59,10 @@ inline size_t CpuTotalPhysicalMemory() { ...@@ -59,9 +59,10 @@ inline size_t CpuTotalPhysicalMemory() {
size_t CpuMaxAllocSize() { size_t CpuMaxAllocSize() {
// For distributed systems, it requires configuring and limiting // For distributed systems, it requires configuring and limiting
// the fraction of memory to use. // the fraction of memory to use.
return std::min(static_cast<size_t>(FLAGS_fraction_of_cpu_memory_to_use * return std::min(
static_cast<size_t>(FLAGS_fraction_of_cpu_memory_to_use *
CpuTotalPhysicalMemory()), CpuTotalPhysicalMemory()),
FLAGS_initial_cpu_memory_in_mb * 1 << 20); static_cast<size_t>(FLAGS_initial_cpu_memory_in_mb * 1 << 20));
} }
size_t CpuMinChunkSize() { size_t CpuMinChunkSize() {
......
python/paddle/batch.py
浏览文件 @ fd9b650d
...@@ -15,7 +15,7 @@ ...@@ -15,7 +15,7 @@
__all__ = ['batch'] __all__ = ['batch']
def batch(reader, batch_size, drop_last=False): def batch(reader, batch_size, drop_last=True):
""" """
Create a batched reader. Create a batched reader.
......
python/paddle/fluid/framework.py
浏览文件 @ fd9b650d
...@@ -382,7 +382,7 @@ class Operator(object): ...@@ -382,7 +382,7 @@ class Operator(object):
'rnn_memory_helper_grad', 'conditional_block', 'while', 'send', 'recv', 'rnn_memory_helper_grad', 'conditional_block', 'while', 'send', 'recv',
'listen_and_serv', 'parallel_do', 'save_combine', 'load_combine', 'listen_and_serv', 'parallel_do', 'save_combine', 'load_combine',
'ncclInit', 'channel_create', 'channel_close', 'channel_send', 'ncclInit', 'channel_create', 'channel_close', 'channel_send',
'channel_recv', 'select' 'channel_recv', 'select', 'gen_nccl_id'
} }
def __init__(self, def __init__(self,
......
python/paddle/fluid/layers/nn.py
浏览文件 @ fd9b650d
...@@ -261,9 +261,10 @@ def embedding(input, ...@@ -261,9 +261,10 @@ def embedding(input,
return tmp return tmp
# TODO(qijun): expose H0 and C0
def dynamic_lstm(input, def dynamic_lstm(input,
size, size,
h_0=None,
c_0=None,
param_attr=None, param_attr=None,
bias_attr=None, bias_attr=None,
use_peepholes=True, use_peepholes=True,
...@@ -324,6 +325,13 @@ def dynamic_lstm(input, ...@@ -324,6 +325,13 @@ def dynamic_lstm(input,
(T X 4D), where T is the total time steps in this (T X 4D), where T is the total time steps in this
mini-batch, D is the hidden size. mini-batch, D is the hidden size.
size(int): 4 * hidden size. size(int): 4 * hidden size.
h_0(Variable): The initial hidden state is an optional input, default is zero.
This is a tensor with shape (N x D), where N is the
batch size and D is the hidden size.
c_0(Variable): The initial cell state is an optional input, default is zero.
This is a tensor with shape (N x D), where N is the
batch size. `h_0` and `c_0` can be NULL but only at the same time.
param_attr(ParamAttr|None): The parameter attribute for the learnable param_attr(ParamAttr|None): The parameter attribute for the learnable
hidden-hidden weights. hidden-hidden weights.
...@@ -387,12 +395,20 @@ def dynamic_lstm(input, ...@@ -387,12 +395,20 @@ def dynamic_lstm(input,
cell = helper.create_tmp_variable(dtype) cell = helper.create_tmp_variable(dtype)
batch_gate = helper.create_tmp_variable(dtype) batch_gate = helper.create_tmp_variable(dtype)
batch_cell_pre_act = helper.create_tmp_variable(dtype) batch_cell_pre_act = helper.create_tmp_variable(dtype)
inputs = {'Input': input, 'Weight': weight, 'Bias': bias}
batch_size = input.shape[0]
if h_0:
assert h_0.shape == (batch_size, size), \
'The shape of h0 should be (batch_size, %d)' % size
inputs['H0'] = h_0
if c_0:
assert c_0.shape == (batch_size, size), \
'The shape of c0 should be (batch_size, %d)' % size
inputs['C0'] = c_0
helper.append_op( helper.append_op(
type='lstm', type='lstm',
inputs={'Input': input, inputs=inputs,
'Weight': weight,
'Bias': bias},
outputs={ outputs={
'Hidden': hidden, 'Hidden': hidden,
'Cell': cell, 'Cell': cell,
...@@ -677,11 +693,13 @@ def dynamic_gru(input, ...@@ -677,11 +693,13 @@ def dynamic_gru(input,
attr=helper.param_attr, shape=[size, 3 * size], dtype=dtype) attr=helper.param_attr, shape=[size, 3 * size], dtype=dtype)
bias = helper.create_parameter( bias = helper.create_parameter(
attr=helper.bias_attr, shape=[1, 3 * size], dtype=dtype, is_bias=True) attr=helper.bias_attr, shape=[1, 3 * size], dtype=dtype, is_bias=True)
batch_size = input.shape[0]
inputs = {'Input': input, 'Weight': weight, 'Bias': bias} inputs = {'Input': input, 'Weight': weight, 'Bias': bias}
if h_0 != None: if h_0 != None:
assert h_0.shape == ( assert h_0.shape == (
size, size), 'The shape of h0 should be(%d, %d)' % (size, size) batch_size, size
inputs['h0'] = h_0 ), 'The shape of h0 should be(batch_size, %d)' % size
inputs['H0'] = h_0
hidden = helper.create_tmp_variable(dtype) hidden = helper.create_tmp_variable(dtype)
batch_gate = helper.create_tmp_variable(dtype) batch_gate = helper.create_tmp_variable(dtype)
......
python/paddle/fluid/tests/book/high-level-api/image_classification/test_image_classification_resnet.py
浏览文件 @ fd9b650d
...@@ -96,10 +96,11 @@ def train(use_cuda, train_program, params_dirname): ...@@ -96,10 +96,11 @@ def train(use_cuda, train_program, params_dirname):
train_reader = paddle.batch( train_reader = paddle.batch(
paddle.reader.shuffle( paddle.reader.shuffle(
cifar10_small_test_set.train10(batch_size=10), buf_size=128 * 10), cifar10_small_test_set.train10(batch_size=10), buf_size=128 * 10),
batch_size=BATCH_SIZE) batch_size=BATCH_SIZE,
drop_last=False)
test_reader = paddle.batch( test_reader = paddle.batch(
paddle.dataset.cifar.test10(), batch_size=BATCH_SIZE) paddle.dataset.cifar.test10(), batch_size=BATCH_SIZE, drop_last=False)
def event_handler(event): def event_handler(event):
if isinstance(event, fluid.EndStepEvent): if isinstance(event, fluid.EndStepEvent):
......
python/paddle/fluid/tests/book/high-level-api/image_classification/test_image_classification_vgg.py
浏览文件 @ fd9b650d
...@@ -73,10 +73,11 @@ def train(use_cuda, train_program, params_dirname): ...@@ -73,10 +73,11 @@ def train(use_cuda, train_program, params_dirname):
train_reader = paddle.batch( train_reader = paddle.batch(
paddle.reader.shuffle( paddle.reader.shuffle(
cifar10_small_test_set.train10(batch_size=10), buf_size=128 * 10), cifar10_small_test_set.train10(batch_size=10), buf_size=128 * 10),
batch_size=BATCH_SIZE) batch_size=BATCH_SIZE,
drop_last=False)
test_reader = paddle.batch( test_reader = paddle.batch(
paddle.dataset.cifar.test10(), batch_size=BATCH_SIZE) paddle.dataset.cifar.test10(), batch_size=BATCH_SIZE, drop_last=False)
def event_handler(event): def event_handler(event):
if isinstance(event, fluid.EndStepEvent): if isinstance(event, fluid.EndStepEvent):
......
python/paddle/fluid/tests/book/high-level-api/understand_sentiment/test_understand_sentiment_stacked_lstm.py
浏览文件 @ fd9b650d
...@@ -87,7 +87,9 @@ def train(use_cuda, train_program, params_dirname): ...@@ -87,7 +87,9 @@ def train(use_cuda, train_program, params_dirname):
def event_handler(event): def event_handler(event):
if isinstance(event, fluid.EndEpochEvent): if isinstance(event, fluid.EndEpochEvent):
test_reader = paddle.batch( test_reader = paddle.batch(
paddle.dataset.imdb.test(word_dict), batch_size=BATCH_SIZE) paddle.dataset.imdb.test(word_dict),
batch_size=BATCH_SIZE,
drop_last=False)
avg_cost, acc = trainer.test( avg_cost, acc = trainer.test(
reader=test_reader, feed_order=['words', 'label']) reader=test_reader, feed_order=['words', 'label'])
...@@ -113,7 +115,8 @@ def train(use_cuda, train_program, params_dirname): ...@@ -113,7 +115,8 @@ def train(use_cuda, train_program, params_dirname):
train_reader = paddle.batch( train_reader = paddle.batch(
paddle.reader.shuffle( paddle.reader.shuffle(
paddle.dataset.imdb.train(word_dict), buf_size=25000), paddle.dataset.imdb.train(word_dict), buf_size=25000),
batch_size=BATCH_SIZE) batch_size=BATCH_SIZE,
drop_last=False)
trainer.train( trainer.train(
num_epochs=1, num_epochs=1,
......
python/paddle/fluid/tests/book_memory_optimization/test_memopt_fit_a_line.py
浏览文件 @ fd9b650d
...@@ -56,7 +56,7 @@ BATCH_SIZE = 200 ...@@ -56,7 +56,7 @@ BATCH_SIZE = 200
# fix the order of training data # fix the order of training data
train_reader = paddle.batch( train_reader = paddle.batch(
paddle.dataset.uci_housing.train(), batch_size=BATCH_SIZE) paddle.dataset.uci_housing.train(), batch_size=BATCH_SIZE, drop_last=False)
# train_reader = paddle.batch( # train_reader = paddle.batch(
# paddle.reader.shuffle( # paddle.reader.shuffle(
......
python/paddle/fluid/tests/unittests/test_merge_ids_op.py 0 → 100644
浏览文件 @ fd9b650d
# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import unittest
import numpy as np
from op_test import OpTest
class TestMergeIdsOp(OpTest):
def setUp(self):
self.op_type = "merge_ids"
ids = np.array([[0], [2], [2], [3], [5], [5], [6]]).astype('int64')
x0 = np.array([[0.1, 0.2], [0.2, 0.3], [0.3, 0.4]]).astype('float32')
x1 = np.array([]).astype('float32')
x2 = np.array([[0.4, 0.5], [0.4, 0.5], [0.5, 0.6],
[0.5, 0.6]]).astype('float32')
out = np.array([[0.1, 0.2], [0.4, 0.5], [0.4, 0.5], [0.2, 0.3],
[0.5, 0.6], [0.5, 0.6], [0.3, 0.4]]).astype('float32')
self.inputs = {'Ids': ids, "X": [('x0', x0), ('x1', x1), ('x2', x2)]}
self.outputs = {'Out': out}
def test_check_output(self):
self.check_output()
if __name__ == '__main__':
unittest.main()
python/paddle/fluid/transpiler/distribute_transpiler.py
浏览文件 @ fd9b650d
...@@ -629,7 +629,7 @@ class DistributeTranspiler: ...@@ -629,7 +629,7 @@ class DistributeTranspiler:
if op.type == LOOKUP_TABLE_TYPE: if op.type == LOOKUP_TABLE_TYPE:
continue_search_lookup_table_op = True continue_search_lookup_table_op = True
op_index = list(all_ops).index(op) lookup_table_op_index = list(all_ops).index(op)
ids_name = op.input("Ids") ids_name = op.input("Ids")
out_name = op.output("Out") out_name = op.output("Out")
...@@ -649,7 +649,7 @@ class DistributeTranspiler: ...@@ -649,7 +649,7 @@ class DistributeTranspiler:
# insert split_ids_op # insert split_ids_op
program.global_block().insert_op( program.global_block().insert_op(
index=op_index, index=lookup_table_op_index,
type="split_ids", type="split_ids",
inputs={ inputs={
'Ids': [ 'Ids': [
...@@ -661,7 +661,7 @@ class DistributeTranspiler: ...@@ -661,7 +661,7 @@ class DistributeTranspiler:
# insert prefetch_op # insert prefetch_op
program.global_block().insert_op( program.global_block().insert_op(
index=op_index + 1, index=lookup_table_op_index + 1,
type="prefetch", type="prefetch",
inputs={'X': prefetch_input_vars}, inputs={'X': prefetch_input_vars},
outputs={"Out": prefetch_output_vars}, outputs={"Out": prefetch_output_vars},
...@@ -672,16 +672,21 @@ class DistributeTranspiler: ...@@ -672,16 +672,21 @@ class DistributeTranspiler:
# insert concat_op # insert concat_op
program.global_block().insert_op( program.global_block().insert_op(
index=op_index + 2, index=lookup_table_op_index + 2,
type="concat", type="merge_ids",
inputs={'X': prefetch_output_vars}, inputs={
'Ids': [
program.global_block().vars[varname]
for varname in ids_name
],
'X': prefetch_output_vars
},
outputs={ outputs={
"Out": [ "Out": [
program.global_block().vars[varname] program.global_block().vars[varname]
for varname in out_name for varname in out_name
] ]
}, })
attrs={"axis": 0})
# delete lookup_table_op # delete lookup_table_op
delete_ops(program.global_block(), [op]) delete_ops(program.global_block(), [op])
......
python/paddle/trainer_config_helpers/attrs.py
浏览文件 @ fd9b650d
...@@ -240,14 +240,15 @@ class ExtraLayerAttribute(object): ...@@ -240,14 +240,15 @@ class ExtraLayerAttribute(object):
:type error_clipping_threshold: float :type error_clipping_threshold: float
:param drop_rate: Dropout rate. Dropout will create a mask on layer output. :param drop_rate: Dropout rate. Dropout will create a mask on layer output.
The dropout rate is the zero rate of this mask. The The dropout rate is the zero rate of this mask. The
details of what dropout is please refer to `here details of what dropout is please refer to `JMLRdropout
<https://www.cs.toronto.edu/~hinton/absps/ <https://www.cs.toronto.edu/~hinton/absps/JMLRdropout.pdf
JMLRdropout.pdf>`_. >`_.
:type drop_rate: float :type drop_rate: float
:param device: device ID of layer. device=-1, use CPU. device>=0, use GPU. :param device: device ID of layer. device=-1, use CPU. device>=0, use GPU.
The details allocation in parallel_nn please refer to `here The details allocation in parallel_nn please refer to `use_case
<http://www.paddlepaddle.org/doc/ui/cmd_argument/ <https://github.com/PaddlePaddle/Paddle/blob/develop/doc/v2
use_case.html#case-2-specify-layers-in-different-devices>`_. /howto/cmd_parameter/use_case_en.md#case-2-specify-layers-in
-different-devices>`_.
:type device: int :type device: int
""" """
......
python/paddle/trainer_config_helpers/layers.py
浏览文件 @ fd9b650d
...@@ -2556,7 +2556,7 @@ def img_conv_layer(input, ...@@ -2556,7 +2556,7 @@ def img_conv_layer(input,
the output will be obtained by concatenating the two results. the output will be obtained by concatenating the two results.
The details of grouped convolution, please refer to: The details of grouped convolution, please refer to:
`ImageNet Classification with Deep Convolutional Neural Networks `ImageNet Classification With Deep Convolutional Neural Networks
<http://www.cs.toronto.edu/~kriz/imagenet_classification_with_deep_convolutional.pdf>`_ <http://www.cs.toronto.edu/~kriz/imagenet_classification_with_deep_convolutional.pdf>`_
The example usage is: The example usage is:
...@@ -5678,8 +5678,8 @@ def warp_ctc_layer(input, ...@@ -5678,8 +5678,8 @@ def warp_ctc_layer(input,
<https://github.com/baidu-research/warp-ctc>`_ library, which is used in <https://github.com/baidu-research/warp-ctc>`_ library, which is used in
`Deep Speech 2: End-toEnd Speech Recognition in English and Mandarin `Deep Speech 2: End-toEnd Speech Recognition in English and Mandarin
<https://arxiv.org/pdf/1512.02595v1.pdf>`_, to compute Connectionist Temporal <https://arxiv.org/pdf/1512.02595v1.pdf>`_, to compute Connectionist Temporal
Classification (CTC) loss. Besides, another `warp-ctc Classification (CTC) loss. Besides, another `warp-ctc repository
<https://github.com/gangliao/warp-ctc>`_ repository, which is forked from <https://github.com/gangliao/warp-ctc>`_ , which is forked from
the official one, is maintained to enable more compiling options. During the the official one, is maintained to enable more compiling options. During the
building process, PaddlePaddle will clone the source codes, build and building process, PaddlePaddle will clone the source codes, build and
install it to :code:`third_party/install/warpctc` directory. install it to :code:`third_party/install/warpctc` directory.
......
python/paddle/v2/minibatch.py
浏览文件 @ fd9b650d
...@@ -15,7 +15,7 @@ ...@@ -15,7 +15,7 @@
__all__ = ['batch'] __all__ = ['batch']
def batch(reader, batch_size, drop_last=False): def batch(reader, batch_size, drop_last=True):
""" """
Create a batched reader. Create a batched reader.
......
Markdown is supported
0% 或 .
You are about to add 0 people to the discussion. Proceed with caution.
先完成此消息的编辑！
想要评论请 注册 或